FAQ

Getting Access

How do I get access to Coffeescribe?

Coffeescribe is in private early access. You need an invite code to create an account. There are two ways to get one:

1. Join the waitlist — Enter your email on the homepage. Codes go out in batches to waitlist members (oldest signups first) as capacity is opened.
2. Get a code from someone — If a friend has a spare invite code (in CS-XXXX-XXXX format), you can use it at signup.

Once you have a code, go to /signup, fill in your email, password, and the invite code, then verify your email with the 6-digit code sent to your inbox. The invite code is redeemed automatically on verification and tokens are added to your account.

What is an invite code and how do I use it?

An invite code is a short, unique, one-time token in the format CS-XXXX-XXXX (for example, CS-AB34-7KJ9). It is issued specifically to you — it isn't shared with multiple people.

To redeem it: enter it in the Invite code field on /signup, complete the signup form, and verify your email. Once verification succeeds, the code is redeemed: tokens are credited to your new account and you're signed in automatically. A code can only be redeemed once — a second attempt will be rejected with "Invalid or expired invite code."

See Waitlist & Invite Codes for the full walkthrough.

Can I sign up with Google or GitHub?

Not during private early access. New accounts require an email address and an invite code — Google and GitHub OAuth are not available for signup right now.

If you already have an existing Coffeescribe account, you can still log in with Google or GitHub via the /login page. Once open access launches, OAuth signup will return.

How long is the waitlist? When will I get an invite?

There is no fixed timeline. Codes go out in batches as capacity is opened up. There is no way to skip the queue, but early signups are prioritised — the sooner you join, the higher your place. You'll receive an email with your personal invite code when yours is ready; keep an eye on your inbox (and spam folder).

What is the Brew Preview — can I try Coffeescribe before I have an account?

Yes. The Brew Preview widget on the homepage lets you try the product anonymously, with no account required. Type a topic (up to 200 characters), click Brew, and the AI generates a real scribe outline — a title, subtitle, a five-to-six chapter outline, and a hook. Pick one chapter and the AI writes that section with live web-search citations. You then see the real sources used, and can open the section in a reading mode where you can look up words or ask the AI questions about your brew.

You can copy or download the section as Markdown, plain text, or PDF. No signup, no tokens, no credit card.

The daily limit is 2 brews per IP address. After that, the button becomes a waitlist CTA.

See Brew a Preview for the full walkthrough.

What is the hero loop on the homepage?

When you land on the homepage before interacting, the hero area plays a short looping animation narrating the full Coffeescribe arc — from typing a topic through live research sources, cited writing, interactive reading, to a finished scribe. It is the attract state, showing the whole product in about 20 seconds.

The moment you click, type in the input, or tap "Try it yourself", the animation hands off immediately to the live Brew widget in the same spot. There is never a split-attention situation — the loop and the live widget share one space.

What is the research feed in the Brew widget?

After a section brews, the widget shows a research feed — the actual web sources the AI searched and used to write the section, with credibility badges and a source count. Every claim in the written section is backed by a real source you can click and read.

This is the visible difference between Coffeescribe and a generic AI chatbot: the writing is grounded in live web search, and you can see exactly what it found.

What is the "read beat" — what can I do after brewing a section?

After brewing, click "Read it" to open the section in a focused reading view. From there:

• Highlight a single word → a popover appears with a short dictionary definition, generated live by AI.
• Highlight a phrase or sentence → a question box appears. Type a question about the highlighted passage and click Ask. The AI answers based only on the text you brewed — it will not go beyond the passage.
• "Save to notebook" → teased here (grayed/waitlisted). Notebooks are an authenticated feature; join the waitlist to claim your brew and access notebooks when you sign up.

All of this is anonymous — no account needed. The define and Ask AI features require that you have already brewed a section (they do not work before a brew). They are also subject to small per-session caps (define: up to 30 lookups; ask: up to 10 questions) and are bounded by the upstream 2-brews-per-day-per-IP limit. No tokens are charged to you.

Can I ask the AI any question in the read beat, or only about the brewed section?

Only about the brewed section. The Ask AI feature in the read beat is deliberately constrained: the answer is grounded in the passage you highlighted and the section you brewed. If your question is not covered by that passage, the AI will say so rather than pulling in outside information.

This is intentional — the read beat is a taste of Coffeescribe's interactive reading experience, not a general-purpose chatbot. For open-ended AI questions, web-search-grounded research, and full document Q&A, those features are available inside the product for account holders.

Will my brewed preview be saved if I join the waitlist?

Yes. If you enter your email in the waitlist form after brewing a section, your brew (the topic, the full scribe structure, and the written section) is saved and pinned to that email address. When you later receive an invite code and sign up with the same verified email, the brew is automatically converted into a real draft scribe in your library under "Your scribes" — the outline becomes chapters and the written section is already there.

The brew is only saved if you choose to join the waitlist. If you close the page without submitting the form, nothing is stored. If you sign up with a different email, the claim does not happen automatically.

What does it mean that my brew has "live web-search citations"?

When the AI writes the section, it runs a live web search (via OpenRouter's web-search plugin) to find real current sources for the topic. The written section ends with a ## Sources list of real links — not fabricated references. This is the core differentiator: it is grounded content, not a stateless prediction from training data alone.

Source links are shown blue and underlined in light mode so they read as real links. Click any to open the original source in a new tab.

The same citation approach powers the full scribe creation experience for members, where every section can be generated with web search enabled.

---

General

What is Coffeescribe?

Coffeescribe is an AI-powered scribe creation platform. You provide a topic and preferences, and AI generates a complete scribe with chapters and sections. You can read, edit, export, and share your scribes.

How does AI scribe generation work?

You enter a topic and use the tile-based form to configure your scribe (Scribe Type, Chapters, Sections, Style, Tone, and more). The AI first creates a scribe structure (chapters and sections), then writes the content for each section. The process is automated but you can control it through different creation modes (Instant Coffee, Americano, Cold Brew). You can also use AI Quick Fill to auto-suggest any fields you haven't filled in — it only fills empty fields and preserves your selections.

What is the Barista?

The Barista is an AI creation assistant available to every signed-in user. Instead of filling in a form, you describe what you want to create in a conversation. The AI asks follow-up questions, determines the best creation mode (Espresso, Americano, or Cold Brew), and builds your brief. You can type or use voice input (Chrome/Edge). When you're happy with the brief, the AI fills in any remaining details automatically and takes you to the form for final review.

What are Custom Instructions?

Custom instructions let you add extra context to your scribe that doesn't fit into the standard form fields. On any creation form, click "Add instruction" in the Final Settings section. Each instruction has a label (e.g., "Format", "Avoid", "Voice") and a description. These are fed directly into the AI generation prompts, giving you precise control over the output. You can add up to 10 per scribe.

Custom Instructions are a different, additive layer from Writing Prompts (Settings → Writing Prompts, or a scribe's own "Edit writing prompt" button) — see What's the difference between Custom Instructions and a Writing Prompt? below.

Is the content I create original?

AI-generated content is created on demand based on your specific topic and instructions. Each scribe is unique. However, AI models are trained on existing text, so outputs may reflect patterns from training data. Similar prompts from different users may produce similar content.

Can I use my scribes commercially?

Yes. Content generated on Coffeescribe is provided to you under a license that allows commercial use, modification, publishing, and distribution. You are responsible for reviewing the content and ensuring it doesn't infringe on existing copyrights.

How accurate is AI-generated content?

AI-generated content may contain errors, inaccuracies, or fabricated information ("hallucinations"). Always verify important facts from authoritative sources, especially for professional, medical, legal, or financial topics. Enabling web search helps improve accuracy for factual content. Scribes generated with web search show a "Web Searched" badge in the library.

Can I fact-check my scribe content?

Yes. Every signed-in user can use Verify Validity in the Workspace to AI fact-check sections or selected text. It extracts claims, searches the web for evidence, and rates each claim as verified, disputed, unverified, or opinion. You can also run a separate plagiarism check. See Verify Validity for details.

Why don't I see all the creation options on the Create page?

All creation modes (AI Create, Guided Creation, Empty Mug, and Start with Research first) are available to every signed-in user. If you can't see them, make sure you're logged in. Your token balance is the only limit — when tokens run out, you'll be prompted to buy a token pack.

What happened to "Espresso", "Americano", and "Cold Brew" on the Create page?

The Create page now uses plain-language labels for first-time clarity: AI Create (form or Barista) covers what was Americano/Quick Book, Guided Creation covers Cold Brew, and Empty Mug is a blank scribe. The original coffee names still live in chat summaries, the Workspace, and the underlying modes — only the entry-point picker changed. You can read the mapping in Understanding Scribe Types.

How do I start with research before creating a scribe?

On the Create page, below the picker cards, click the "Start with Research first" band. This takes you to Research Mode where you can build a Cafe — a workspace for collecting web searches, academic papers, YouTube transcripts, scraped pages, and uploaded documents. Once your Cafe has the material you want, come back to /create and pick how to create your scribe; you can reference your research notes from the Workspace using @ mentions. Research Mode is available to every signed-in user.

How do I create a blank scribe with no AI generation?

Pick Empty Mug on the Create page (available to every signed-in user). It reveals an inline title field on the same page — type a name and click Create. You're taken straight to the Workspace with a single blank chapter and section, ready to edit manually or fill in with AI Write later.

Why is the create form a slideshow now?

As of May 2026 (E64.3), both AI Create ("Fill out a form") and Guided Creation ("Fill out a form") use a 6-card slideshow instead of a single long form: Title → Topic & Purpose → Audience & Reader Level → Structure → Description → Review. One focused question at a time, with a Review at the end where every answer appears as a small tappable card. The slideshow keeps the form short and lets you skip anything you're unsure about.

What's the difference between Quick Read and Full Scribe?

On Card 1 — Title & length of the create slideshow, the Target length tile group lets you pick:

• 10 / 20 / 30 / 60 min — a Quick Read (Espresso) — a condensed scribe with a fixed total word budget of 250 wpm × N min (so a 20-min Quick Read aims for ~5,000 words across however many sections fit).
• Full scribe — a full-length book — multi-chapter Americano or Cold Brew with up to 24 chapters and no total word cap.

In both cases the AI now decides variable section counts per chapter from your title, audience, style, and tone — not a uniform N × M shape. After submission, AI Create and Full Scribe drop you on a Structure Review page where you can rename, add, and delete chapters and sections before writing begins. See Structure Review for details.

Note (E67, May 2026): The Target length tile group moved from Card 4 to Card 1 alongside the title, so AI Quick Fill knows the depth and tone to aim for before it runs.

What is the Structure Review screen?

After you click Create Scribe on AI Create or Full Scribe, you land on /book/[bookId]/structure-review — a dedicated page where the AI has proposed a chapter-and-section layout. You can rename, add, or delete chapters and sections, then click Proceed to Writing to start prose generation. The page is refresh-survivable — close the tab, come back later, and your edits are still there. See Structure Review for the full walkthrough.

Empty Mug skips Structure Review (there's nothing to review — you start with a blank chapter). Guided Creation has its own equivalent screen at /create/guided/[bookId] that uses the same editor.

Why do my chapters have different numbers of sections now?

Before May 2026 (E67), every scribe had a uniform N × M shape — e.g. always 10 chapters × 4 sections. The result was often awkward: a small topic got padded out, a big topic got cut short. Now the AI picks variable sections per chapter from your title, length, audience, and style — Chapter 1 might have 2 sections; Chapter 4 might have 5. Quick Read totals stay within the duration's word budget; Full Scribe stays under 24 chapters. If you want a uniform shape, you can edit on the Structure Review screen, or use the legacy Chapters/Sections tiles on Card 4 (Guided + Full Scribe) as overrides.

What is Simple Mode?

Simple Mode is the quickest way to a scribe — just a topic and a length tile. It lives at /brew and there's an amber Simple Mode entry button on the right side of the secondary band on /create. No multi-card form, no slideshow — type a topic, pick 10 / 20 / 30 / 60 min or Full scribe, and submit. You'll still go through the Structure Review pause for AI Create-style flows. Available to every signed-in user.

What's the difference between Simple Mode and AI Create?

Both are AI-driven. Simple Mode at /brew is a one-form flow — topic, length, submit, done. AI Create at /create (the leftmost card) opens the 6-card slideshow with optional Audience / Style / Tone / Description fields and access to AI Quick Fill, the model picker, custom instructions, and so on. Use Simple Mode when you just want output fast; use AI Create when you want to shape the brief before generation.

Why does structure generation feel slower now?

Structure generation (the step that picks chapter count and variable section counts) runs on Sonar (Perplexity) for every signed-in user since E67. Sonar is a research-grounded model with built-in web search — slower than a generic chat model, but it produces structures that reflect what's actually written about your topic on the web rather than the model's stale training data. Section writing uses your selected model (the CoffeeScribe Model by default, or any model you pick from the picker), so per-section speed is unchanged. The trade-off is most noticeable on a 10-min Quick Read where the structure step is a larger fraction of total time.

What happens if I leave fields blank in the create form?

Skipping is fine — the hint at the top of every card reminds you: "Skip anything you're not sure about — AI can fill the rest on Review." On the Review card you'll see two ways to fill blanks:

• ✨ Let AI fill the rest — populates only the fields you skipped, preserving anything you typed. You can edit or delete what AI suggests before generating.
• Sticky purple FAB ("AI fill everything") at the bottom-right of every card — fills every blank field and jumps you straight to Review.

If you click Create Scribe with optional fields still blank, a confirm modal appears with three options: Let AI fill the rest first, Create with what I've given (AI fills any gaps with sensible defaults during generation), or Cancel to keep editing.

Can I save a draft of the create form and come back later?

Yes. Your slideshow progress saves automatically to your browser as you go. If you close the tab and return to /create/quick or /create/guided, you'll see a "Resume your scribe?" prompt with [Resume] (pick up where you left off) or [Start fresh] (clear and begin again). Drafts are stored locally on your device — they aren't synced across browsers — and expire after 7 days.

What is "AI fill the rest"?

On the Review card of the create slideshow, ✨ Let AI fill the rest fills only the fields you skipped (preserving what you typed). It uses your title and topic as context to suggest sensible values for description, goals, audience details, and so on. Nothing generates silently — you can review, edit, or delete the AI's suggestions before clicking Create Scribe.

This is different from the sticky purple FAB ("AI fill everything") which appears on every card and fills every blank in one shot before taking you to Review.

Why is the button now "Create Scribe" instead of "Create Book"?

Coffeescribe's product naming uses Scribe for the unit of work — the thing you create, read, share, and export. The button on the create slideshow Review card is now Create Scribe to match the rest of the product (Scribe Hub, Audioscribe, Empty Mug, "Brewing Your Scribe"). The behaviour is unchanged — it kicks off generation and takes you to the live progress page.

Where did Advanced settings (model, generation mode, custom instructions) go?

On the create slideshow they live in an Advanced settings accordion on the Review card, below the Generate button. The accordion is collapsed by default to keep the default surface clean. Expand it to access AI model selection, web search toggle, generation mode (Lightning / Pour Over / Slow Brew), language, custom focus prompt, and custom instructions. Every signed-in user gets the full model picker — the CoffeeScribe Model (the curated default we maintain) is pre-selected, but you're free to pick any other model. Web search, generation mode, language, custom focus, and custom instructions are available to every signed-in user in the accordion.

Can I make the Advanced settings stay open by default?

Yes. Go to Settings → Preferences and turn on "Show advanced controls". After that, the Advanced settings accordion on the create slideshow Review card opens automatically every time you reach Review — so web search, generation mode, custom instructions, and the model picker are visible without an extra click. Toggle it off and the accordion goes back to being collapsed by default. The preference is per-account and persists across browsers and devices (it's saved to your account, not just your browser), and works the same way for every signed-in user.

---

Importing your own books (Scribe Conversion)

What is Scribe Conversion?

Scribe Conversion lets you upload a document you own — PDF, EPUB, Word, plain text, or Markdown — and turn it into a native scribe in your library. From /create, click Upload your own. You then pick one of three conversion modes: Word-for-Word (faithful, cheapest), AI Rewrite (rewrites the whole book in a style you choose), or Mini-Scribe (condensed version). See Importing your own books for the full walkthrough.

What file types can I import? And how big?

PDF, EPUB, DOCX, TXT, and MD up to 100 MB. Scanned PDFs are OCR'd automatically (Mistral OCR, billed against your token allowance — typically $0.30–$1.40 per book). DRM-protected EPUBs are detected and rejected. Files with mismatched extensions (e.g. a PDF renamed to .epub) are also rejected — we check the actual file signature.

Does Scribe Conversion cost tokens?

Yes. Each mode shows an estimated token cost before you commit:

• Word-for-Word — cheapest, just structure detection.
• AI Rewrite — proportional to word count × your selected model.
• Mini-Scribe — proportional to word count, generally cheaper than AI Rewrite.
• OCR — only for scanned PDFs, billed at upload time, $0.30–$1.40 per book.

If you don't have enough tokens, the convert button is disabled with a top-up link.

Who can use Scribe Conversion?

Every signed-in user. It is not tier-gated. The only limit is your token balance.

Are my scribes public by default?

No — every scribe is private by default, for every signed-in user. Only you can see a new scribe until you choose to publish it. Right after creation you'll see a "Your book is brewing!" dialog with a Private / Public toggle (it defaults to Private), and you can change a scribe's visibility any time from the scribe's page. Publishing puts it in the public library where signed-in members can read it; switching back to private removes it.

Are imported scribes public by default?

No — imported scribes default to private. They live in your library and are not visible to anyone else until you explicitly publish them via the Publish to library flow. Publishing requires a separate consent confirmation in which you re-affirm you hold the rights to share the work publicly.

What if I publish a scribe I don't own the rights to?

We will remove it from public view. Coffeescribe's stance is direct: if you did not write it, and AI did not generate it, then you uploaded it. It could be plagiarism. Even if AI then rewrote it, it could still be plagiarism. We have a full right — and a responsibility — to remove that content from public view. AI rewriting does not launder copyright. An AI Rewrite or Mini-Scribe derivative of a copyrighted source is still derivative of that source, and you remain responsible for the rights upstream. If a rights-holder reports a published imported scribe, we remove it and notify you. Repeat infringement may result in account suspension or termination. Please only publish content you have the rights to — and aim for powerful content. See Terms of Service section 3.5 for the full text.

Are images and figures preserved in imported scribes?

Not in v1 — only text comes through. You'll see a warning in the dropzone and again on the structure-review screen. If you need images, treat the import as a starting point and re-add them in the Workspace.

What is the "Manual split" tab on the review screen?

It's grayed out in v1 with "Coming in V2 — drag-to-highlight section markers on PDF preview." For now, pick AI-proposed and edit the chapter list in the Workspace after conversion, or pick Use as-is if the source already has clean chapter breaks.

Why do scanned PDFs cost more?

Two reasons: (1) Mistral OCR charges per page of text it extracts, and (2) the v1 pipeline re-runs OCR during the convert step instead of caching the parsed text, so a scanned PDF roughly doubles in cost compared to a text PDF of the same length. We plan to fix the second part in a follow-up release.

Can I leave my import running and come back later?

Yes. The whole pipeline is resume-safe. You can close the tab during upload, OCR, or generation — the server keeps working. To pick up where you left off, go back to /create/import?resume=<uploadId> (the URL is preserved in the address bar while you're in the flow) or check your library when it finishes.

What's the "Convert to Scribe" button in the Workspace?

If a scribe grows too large for the Workspace editor to handle smoothly, you'll see a too-big banner with a Convert to Scribe button. It pre-fills the import flow with the current scribe as the source — useful when you want to re-shape a sprawling work into a cleaner Mini-Scribe or AI Rewrite.

Why don't my imported scribes show "Chapter N:" in the title?

Most published books already include the chapter number in the title (e.g. "Chapter 1: The Beginning"). On imports, CoffeeScribe drops the synthetic "Chapter N:" prefix in the viewer and Read view so the title displays as the author wrote it — no duplication.

How do I find my imported scribes in the library?

Use the Imports filter chip in the library. Imports also sort newest-first by default so a fresh upload is easy to spot at the top of the list.

---

Subscription & Pricing

Coffeescribe is currently pay-as-you-go: any signed-in user buys one-time token packs ($10 → 10M tokens, $20 → 20M, $50 → 50M) through Polar, our Merchant of Record. There's no subscription to buy right now — see Plans & Pricing and the Billing FAQ for the full picture. Every signed-in user reaches every feature — the full model picker, audiobooks, all export formats, all Research Mode tools — gated only by your token balance. Free, Pro, and Creator still exist internally as account labels (a holdover from an earlier subscription model), but no feature or price depends on which one your account has.

What do new users get?

Every account can access the full app — every creation mode (AI Create, Guided Creation, Empty Mug), Workspace, Research Cafes, web search, all export formats, the full 350+ model picker, and audiobook generation. Scribes are private by default for every account (see Are my scribes public by default?). When your tokens run out, buy a token pack to keep going, or bring your own AI provider key (BYOK).

What's included?

See our Plans & Pricing page for the full feature list — it's the same list for every signed-in user. Token packs don't unlock anything extra; they just add to your balance.

How do tokens work?

Tokens are the units AI models use to process text. You get tokens by buying token packs (or from an invite code, if you were given one); different models consume tokens at different rates. See Understanding Tokens for details.

Where's the fastest way to buy tokens?

The Tokens button in the header — it's visible on every page once you're signed in and opens a Buy Tokens dialog with the same token-pack grid as the Pricing page. Pick a pack and you're taken straight to checkout.

Where do I see my token balance?

Your exact token count is shown in Settings → Billing and in the account menu (click your avatar) — the real number, not a vague meter, so you always know exactly what you have before you spend it.

How many PDF downloads do I get?

Unlimited, for every signed-in user, at no token cost — the same is true of Word, EPUB, and StoryBook scribe exports, and of research report PDFs (Auto-Write, Verify, Conflict, and Find the Truth reports, W54).

What happens when my tokens run out?

When your balance reaches zero, any action that costs tokens (creating scribes, research, AI analysis) shows a prompt to buy a token pack. Your existing scribes, highlights, notes, and exports remain accessible — you just can't generate new content until you top up (or use BYOK).

Can I change my plan?

There's no subscription plan to upgrade, downgrade, pause, or cancel right now — Coffeescribe runs on one-time token packs, and Settings has no billing-cycle management to do. Every signed-in user already has full feature access, so there's nothing to move between. See Buying Tokens & Plan Changes.

What happens if I run out of tokens?

Any action that costs tokens shows a prompt to buy a token pack — this works the same way for every account. A bigger balance always comes from buying a pack (or BYOK).

Partially generated scribes are saved and can be completed later — see below.

What if my tokens run out partway through a scribe being generated?

The token balance is checked before each section is written, not before the whole scribe is started. If you start a long scribe and your balance hits zero halfway through, generation stops cleanly at the next section — you don't get charged for a section that never runs. The chapters and sections written up to that point are saved, and the scribe is marked partial. Buy a token pack and resume — the system picks up from where it stopped.

This applies to anyone with a finite balance. A very long scribe (10+ chapters with many sections each) can still exhaust a small balance before completion, especially on a pricier model.

Do unused tokens roll over?

There's nothing to roll over — purchased token packs never expire. Every token you buy stays in your balance until you spend it, however long that takes.

Can I get a refund?

Refunds are handled through Polar, our Merchant of Record. Tokens already spent on completed generations are non-refundable. See the Billing FAQ and Refund Policy for full details.

---

Content & Quality

Why do different AI models produce different results?

Each AI model is built and trained differently. Every recommended model is scored on Writing, Research, and Cost (1-5 dots) so you can compare at a glance. Budget models are faster and cheaper but may produce simpler writing. Premium models tend to produce more nuanced, polished content. Tap the info icon next to any model to see its strengths and weaknesses.

How do I improve content quality?

• Be specific with your topic and instructions
• Use a premium AI model
• Enable web search for factual content
• Use the Research panel to gather information and save it as notes
• Reference research notes in AI Write using @ references for grounded, accurate content
• Add per-section instructions to guide the AI on each section
• Edit and refine in the Workspace
• Provide clear guidance on tone, audience, and focus

What are the author attribution options?

When you publish or export a scribe, you choose how your name appears: Created by (AI-generated with minimal editing), AI-Assisted by (you significantly edited the AI content), or Written by (100% human-written). If your scribe has a mix of AI and human-written sections, "AI-Assisted" is suggested automatically. You can change the attribution anytime from the export or attribution modal.

What's the difference between "Created by" and "AI-Assisted by"?

"Created by" means the content is primarily AI-generated. "AI-Assisted by" means you made substantial edits or rewrote portions of the AI output. Use whichever best describes your involvement. "Written by" is reserved for content you wrote entirely yourself — a compliance warning is shown if you select it.

Can I edit the AI-generated content?

Yes. Every signed-in user has full access to the Workspace where you can manually edit text, rewrite sections with AI, ask AI questions, research topics with web search, upload reference documents, and restructure chapters. The section editor is WYSIWYG — type normally, use the formatting toolbar (Bold, Italic, headings, lists, tables, etc.) or keyboard shortcuts (Ctrl+B, Ctrl+I, Ctrl+Z) to format. No Markdown knowledge required. Every signed-in user can also export to Word for external editing.

What's the difference between the generation modes?

Coffeescribe offers four generation modes from fastest to finest: Lightning (all sections in parallel, cheapest), Lightning Medium (chapters parallel, sections sequential within each), Pour Over (chapters sequential, sections parallel within each), and Pour Over Slow Brew (every section one by one, maximum coherence). The token cost gap grows with more chapters and pricier models. Use Lightning for knowledge scribes and quick drafts; Pour Over or Slow Brew for creative writing and narratives. See Generation Modes for details.

Why did my scribe generate in Pour Over Slow Brew (or a slower mode) instead of Lightning?

As of July 2026, Generation Mode pre-selects a sensible default based on your Genre and Scribe Type — you're not stuck with one universal default anymore. Fiction and Comedy genres, and any story-shaped scribe (story, essay, article, blog post, or training course — memoir included), default to Pour Over Slow Brew because narrative writing needs every section to remember what came before (character names, plot points, prior arguments). Whitepapers and reports default to Pour Over. Everything else (book, guide, custom, or factual genres) keeps the original Lightning Medium default — never the fully-parallel Lightning mode, since a context-free mode is never a good default. This is only a starting point: open Advanced settings on the create form's Review card and pick a different mode yourself any time — your manual choice always sticks, even if you change the genre afterward. See Generation Modes for what each mode does and its cost trade-offs.

Why is Web Search off by default for my story or comedy scribe?

Web search now defaults based on your Genre and Scribe Type. Fiction is locked off entirely (the toggle is disabled with a tooltip explaining that it protects the writing voice). Comedy and any story-shaped scribe — including memoirs and other factual-but-narrative books — default to off, but you can flip the toggle on any time, for example if your story leans on real people or events. Factual genres (non-fiction, self-help, educational, business, technical, and similar) still default to on. Whichever way you leave the toggle sticks once you've touched it, even if you change the genre afterward. See Understanding Scribe Types for the full matrix.

Why doesn't my story show citations, even though I turned Web Search on?

Story-shaped scribes (story, and any memoir or narrative non-fiction using that Scribe Type) never show citation apparatus — no inline [N] markers, no clickable source links, no "References"/"Sources" block — even with Web Search manually turned on. The research still runs and still informs the writing; it's woven into the prose invisibly instead of being cited, because citation markers interrupt narrative voice. Every non-story scribe type (book, guide, whitepaper, report, article, blog post, training course) keeps normal citation behaviour. See Citations & References.

Does AI Quick Fill know that some form settings affect each other?

Yes. AI Quick Fill and the AI Barista now understand the knock-on effects of the fields they fill — for example, both know that picking Comedy should default Web Search off and Generation Mode to Slow Brew, so they won't hand you a combination that fights itself. They also keep genre words (like "Comedy" or "Memoir") in the Genre field and out of the Scribe Type field, so the scribe resolves the way you'd expect instead of, say, "Comedy" accidentally landing as a custom Scribe Type while Genre stays generic non-fiction.

If my scribe stops mid-generation and I click "Complete Scribe," does it write the rest in order?

For narrative-mode scribes — anything using Pour Over Slow Brew, Pour Over, or Lightning Medium — yes. The remaining sections are generated strictly in book order, one at a time, each threaded with everything written before it as context, exactly as if generation had never stopped. This matters most for a Slow Brew story that got interrupted partway through: without it, the resumed sections could fan out in parallel and lose track of characters or plot established earlier. Fully-parallel Lightning mode is unaffected, since its sections never depended on each other in the first place.

Why does my scribe's writing style change based on reading level, style, and tone?

The writing prompt now calibrates much more consistently to the Reader Level / Age Range, Style, and Tone you picked (or that AI Quick Fill picked for you). A scribe aimed at "Under 12" gets meaningfully simpler vocabulary and shorter sentences than the same topic aimed at "Expert" readers, and a "Casual" style reads differently from "Academic" — including a previously-reported issue where children's-level scribes still read like they were written for adults. These have always been form fields; the improvement is that the AI now reliably applies them.

How do I add research to my writing?

There are several ways to use research in the Workspace:

1. Click Research in the toolbar to search the web and save results as research notes
2. Click Upload to import a PDF, DOCX, TXT, or MD file as a research note
3. In the AI Write cockpit, check research notes in the right panel to include them as context
4. Type @ in any instructions field (AI Write or per-section instructions) to search and reference saved research notes

See Workspace and Research Mode for full details.

What file types can I upload as research?

You can upload PDF, DOCX (Word), TXT (plain text), and MD (Markdown) files. The maximum file size is 5 MB. Uploaded files are saved as research notes and can be referenced when generating content with AI.

My sidebar keeps collapsing when I click a section — what gives?

You probably haven't pinned it. Click the pin icon at the top of the sidebar — a pinned sidebar stays open across every section click, so you can browse chapter-to-chapter without it snapping shut. An unpinned sidebar will auto-collapse on click (to give you maximum room for the editor) and re-open when you hover over the left edge again.

Why does the sidebar sometimes open when I'm just moving my mouse across the screen?

It shouldn't. We added a 150 ms delay before the sidebar opens on hover so a quick cursor pass-by doesn't trigger it. If you're still seeing it open on transit, let us know — that's a bug.

How do I reorder sections within a chapter?

Hover over the section row in the sidebar — the grip handle appears on the left edge (always visible on touch screens). Grab and drag the section up or down to reorder. The new order saves immediately. Cross-chapter drag-and-drop isn't supported in this release — drop a section onto a different chapter and you'll see a "Cross-chapter moves coming soon" toast. For now, deleting + re-adding is the workaround.

How do I combine two sections into one?

Hover over a section, click the ⋯ menu, and choose either Combine with section above or Combine with section below. The two sections merge into one — the lower-numbered section's title is kept, and the content is joined with a blank line. If you change your mind, click Undo in the toast that appears (you have 10 seconds). Any highlights and notes on the merged section are automatically moved to the surviving section so you don't lose them.

What are Cafes in Research Mode?

Cafes are independent research workspaces where you collect material from multiple sources — web searches, academic papers, YouTube videos, and web pages. Each Cafe has its own Serving Tray for curating the best snippets. You can create multiple Cafes for different research topics. Find them under Research in the navigation bar. Available to every signed-in user — the CoffeeScribe Model is pre-selected by default for in-Cafe AI operations, and every user can pick a different model from the picker.

What are Café Groups, and how do I create one?

As of W1.43, you can organize your cafés into groups on the /research page — think folders. On desktop, drag one café card onto another to create a group; you'll be asked to name it. On any device (including a phone), click a café's ⋮ menu and choose "Add to group…" to pick an existing group or create a new one. Groups are pure organization — they don't cost tokens and don't change anything about the café itself. See Research Mode → Organizing Cafés into Groups for the full walkthrough.

Can a café be in more than one group?

Yes — groups are many-to-many, not exclusive folders. An "Avocados" café can live in Health, Super foods, and Healthy fats at the same time. Dragging a café into a group (or using "Add to group…") always adds it — it never removes the café from anywhere else it already belongs.

If I delete a group, does it delete my cafés too?

No. Deleting a group only removes the group itself — every café that was in it is untouched and still exists on your /research page (find it in another group it belongs to, or in the main grid if that was its only group). The delete confirmation tells you exactly how many cafés will survive before you confirm.

How do I group cafés on my phone?

Drag-and-drop isn't available on a phone — it would fight the page's scroll gesture. Instead, tap a café card's ⋮ menu and choose "Add to group…". This picker does everything drag does on desktop: join an existing group or create a new one.

Can other people see my Café Groups?

No. Groups are private to you — they never appear on a shared café. If you send someone a read-only café link, nothing about that café's group membership (or whether you have groups at all) is visible to them.

Why does the same café show up twice when I'm importing sources or tray items?

Because it's in two (or more) of your groups. The cross-café pickers — importing sources into a café, or importing tray items into a book chapter — organize cafés by group, so a café that belongs to more than one group is listed once under each group it's in.

What academic databases can I search?

You can search three academic databases from the Academic dropdown in the search bar:

• OpenAlex — 250M+ works across all disciplines. Best for broad searches with citation counts. Includes journals and preprints.
• arXiv — Preprints in physics, math, computer science, and related fields. Best for cutting-edge research with direct PDF links. Not peer reviewed.
• PubMed — 35M+ biomedical and life sciences articles from NCBI. Best for medical, biology, and health sciences. Mostly peer reviewed.

All three cost the same — a small portion of your allowance per search — and feed into the same Analyze and Add Raw workflows.

What sources can I add to a Cafe?

You can add these types of sources:

• AI Web Search — AI-powered web search with synthesized responses and citations
• Academic Papers — Search millions of papers via OpenAlex, with optional AI analysis
• YouTube Transcripts — Extract transcripts from any YouTube video (search bar auto-detects the URL)
• URL Scraping — Extract content from any web page or article (search bar auto-detects the URL)
• File Upload (documents, images, audio, video) — As of W1.35 P4, click the paperclip in the Ask Your Cafe chat composer and choose Upload a file. See What file types and sizes can I attach in a café chat? below.
• Notebook Import — Import highlights and notes from your Scribes (the Add sources button in the Cafe's view bar)

As of W1.42, click Add sources in the Cafe's view bar (top of the workspace, next to Share) to open a dialog with three more ways in — Paste text, Import from notebooks, and Import from another cafe (W1.20, which pulls in Serving Tray items from a different Cafe you own rather than gathering fresh — see the next few questions). Uploading a file is not one of them; it lives in the chat composer's paperclip instead. (As of W1.44, the view bar no longer carries a Deep research menu — every research verb, including Auto-Research which gathers sources, now launches from the ✦ Deep button on the chat composer; see What happened to the Deep research menu? below.)

Can I reuse research from another Cafe instead of gathering it again?

Yes — click Add sources ▾ → Import from another cafe in any Cafe. Pick one of your own Cafes, select items from its Serving Tray (or Select all), and click Import. Each selected item lands in the current Cafe as a real source in the Sources rail and its excerpt on the Serving Tray, ready to cite in Auto-Write, Find the Truth, Claims, and Ask Your Cafe. See Research Mode → Import from Another Cafe for the full walkthrough.

Does importing from another Cafe copy or move the research?

Copy, always. The origin Cafe is completely untouched — its Serving Tray and Sources rail keep everything they had. Import only ever adds a duplicate copy to the destination Cafe; there's no way to "move" research out of a Cafe.

Why can't I import from a Cafe someone shared with me, or a public Cafe?

Only Cafes you own appear in the import picker. CoffeeScribe's Cafe-sharing model is a read-only link — it doesn't record who you shared it with, so there's no list of "Cafes shared with me" to draw from yet. Importing from a shared-with-you or public Cafe is planned for a future release; for now, save the specific snippets you want to your own tray from within the shared Cafe first, then import normally.

Will importing duplicate sources I already have?

No. A source is matched against what's already in the destination Cafe (by DOI, PMID, arXiv ID, ISBN, URL, or title — whichever it has) before it's copied. If it's already there, the import skips adding a second copy and attaches the incoming excerpt to the source that's already present instead. The result toast tells you exactly how many were added versus skipped, and re-importing the same tray item twice won't create duplicate excerpts either.

Does importing from another Cafe cost tokens?

Yes, a small amount. Each copied source and copied tray item needs to be embedded in the new Cafe so it's searchable in Ask Your Cafe — that's the same one-time embedding cost as adding any new source (see Embedding Costs). You are not billed again for whatever it originally cost to gather the source in the origin Cafe — only for making the copy usable here.

What is the Serving Tray?

The Serving Tray is the right rail in each Cafe. It holds curated snippets — the most useful bits from your research. Highlight text anywhere and use "Add to tray" to save it. You can also add blank notes with the + button for pasting external content. As of W1.13, the tray defaults to a slim collapsed strip (count + top 3 titles) so it doesn't crowd the workspace — expand it to see the full list, drag to reorder, and click any item to read or edit it in the central reading pane. Combine multiple items into one via the Select toggle.

Where did the source cards go? Why does everything look flatter now?

As of W1.13 (and restructured again in W1.42), the Cafe workspace is three columns instead of expanded cards everywhere: a left panel (Chats · Reports · Sources tabs, compact rows rather than full cards) on the left, a central reading pane where the full detail of whatever you click — a source, a tray item, or a report — actually renders, and a Serving Tray on the right. Nothing shows its full content in the side columns anymore; click a row and it opens in the pane with a single toolbar (Analyze, Verify, Add to tray, Copy, Download, Delete, ⓘ metadata) instead of a 5-button row repeated on every card. This declutters the view at large source counts while keeping every capability the old expanded cards had. As of W1.44, each of the three tabs also shows a live count next to its label (how many chats, how many distinct reports, how many sources) so you can tell what's actually in a tab without opening it.

What happened to the Deep research menu?

It's gone, as of W1.44 — replaced by a ✦ Deep button on the chat composer (last item on the row, right next to Send). Every research verb it used to hold is still there, under the same three headings (Gather, Interrogate, Write); nothing was removed. Click ✦ Deep, pick a verb — or, for Auto-Research, one of its three depth chips right under its row — and its setup fills the centre column where your chat and reports already are, instead of opening as a window over the top of them. If you've typed a question in the chat box first, it carries into Auto-Research's topic or Find the Truth's claim automatically. A run in progress shows a pulse on ✦ and a "● running — view" row in the picker, from anywhere in the café. See Research Mode → Starting a Run: The Deep Launcher for the full walkthrough.

Why did my downloaded source PDF look unformatted, with "Untitled Source" as the title?

That was a known issue, fixed as of W62. Source and extraction-card PDF downloads previously showed "Untitled Source" as the title, rendered tables as raw | … | pipe text, and named the file after the Cafe plus a random ID instead of the card's actual topic. Re-downloading now gives you the card's real title (matching what's on-screen), a matching filename, styled headings/bullets/real tables, and — for a card derived from a YouTube transcript or a scraped web page — a source-attribution line naming the true origin.

Why are my sources grouped by run now?

As of W1.14, the sources rail's top level is an index of the runs that produced your sources — Auto-Research, Auto-Write, Detect Conflicts, Find the Truth, Verify reports, and Individual sources (manually-added, always last) — instead of a flat list with a view-preference grouping knob. Click a run to lazy-load only its sources underneath; the rest of the rail stays light. This replaces the old "Group by Question/Type/Used/Detail/Verify" control and the "Show" narrow-to-type dropdown, both of which are gone — the point is to find "all the sources for the one question I asked" in one click, and to keep a Cafe with hundreds or thousands of sources responsive since nothing loads until you open its run.

Why is "Individual sources" split into YouTube / Links / Uploads / Papers / Other now?

As of W1.25, Individual sources is no longer one undifferentiated bucket — it's split into five sub-groups (YouTube, Links, Uploads, Papers, Other sources), each lazy-loaded the same way a run is. This is only for manually-added sources; anything produced by a run (Auto-Research, Auto-Write, etc.) still lives under that run's own section, unaffected. A source with extraction cards (Summary, Key Points, Fact Check) now shows them nested directly underneath it via a disclosure toggle, and opening one source's extractions closes whichever other source's was open — an accordion, so only one set of extractions is visible at a time.

How do I delete a whole research run or question?

Each run row (Auto-Research, Auto-Write, Detect Conflicts, Find the Truth, Verify) has a trash icon. Click it and confirm in the dialog that appears — this removes that run's report/verdict and its sources from the Cafe in one action, instead of deleting sources one at a time. Individual sources isn't a run and has no whole-group delete; remove those one at a time as usual.

What happens to my saved (tray) sources when I delete a run?

They're kept. Deleting a run only removes what's still living in the sources rail under that run — anything you already pulled into your Serving Tray is a separate, deliberately-saved copy, and a run delete never touches your tray.

Where did the old sources filters go?

The free-text search and the Well-researched / Verified toggles moved from the top of the rail (W1.13) to inside an expanded run (W1.14), as a "Refine this run…" control that filters only that run's already-loaded sources. Expand a run to see it.

How much does Research Mode cost?

All research actions cost tokens. AI Web Search uses the most — a meaningful slice of your allowance per query. Academic paper search, YouTube transcript extraction, URL scraping, and Ask AI / Analyze each cost only a small portion of your allowance. Token estimates are always shown before you confirm an action so you know what you're spending.

What is Analyze source (formerly "Ask AI") on research sources?

Every saved research source has an Analyze source button (sparkle icon). Click it to run AI analysis with presets (Summarize, Key Points, Fact Check, Topics & Structure) or a custom prompt. You can select multiple presets at once. This works on any source type — web search results, transcripts, scraped pages, uploads, or imported notes.

This button used to say "Ask AI" — it was renamed so it doesn't get confused with the new highlight-anywhere Ask AI popover (see below), which is a quick chat about a specific highlighted snippet rather than a full structured analysis of the source.

How do I ask AI about something I'm reading in a Cafe?

Highlight any text — on a source card, a Q&A reply, a tray item, or a research report — and a small popover appears with Ask AI · Verify claim · Save to notebook · Add to tray. You can also right-click a source card to get the same popover scoped to the whole source. Click Ask AI and the "Ask your Cafe" chat opens pre-filled with your selection, ready to send. If you highlighted text inside a specific source or report, the answer automatically takes into account which one it came from.

Can I look up a word while researching?

Yes. Highlight a single word anywhere in a Cafe (a source, a report, the tray) and a Dictionary action appears in the popover instead of the usual four — click it for an instant AI definition, the same popup used in book Reading Mode. There's a "Save to notebook" button right in the definition popup if you want to keep it.

How do I save an Ask-AI answer or a definition to my notebook?

From the highlight popover, click Save to notebook to save a selected passage as a note (grouped under the Cafe's name in Notebooks). From a Dictionary definition, click the Save to notebook button inside the definition popup. From a full Q&A chat answer, use the +Notebook button under the answer.

What's the difference between Grounded and General in "Ask your Cafe"?

As of W1.42, Ask Your Cafe has one mode chip on the left of the question box, reading "Web" or "Cafe sources" — click it to cycle between the two, and a helper line explains whichever is active. ("General ask AI"/"Grounded from cafe sources" are the same two modes' internal names, from before the composer was rebuilt into a single row — you may still see those fuller names elsewhere in this Help Center.) "Web" is the default: it answers from the AI's own knowledge, running on whatever model is currently selected in the model picker, usually with live web search (the toggle defaults to on — see the next question) — and its answers stream in live as they're written. "Cafe sources" is the opt-in: it answers strictly from your in-scope Cafe sources, tray items, and any chats you've added to your cafe memory, with citations — if your research doesn't cover the question, it says so instead of guessing. As of W1.25, Cafe sources is pinned to a fixed model (DeepSeek V4 Flash) and never searches the web, regardless of what's picked elsewhere on the page. As of W1.44, your choice is remembered separately per Cafe (not for your browser as a whole — before W1.44, flipping the chip in one Cafe silently changed the default for every Cafe you own) and persists across reloads (see Which mode am I in by default? for the one exception). If a Cafe-sources answer comes back as a refusal, an "Ask the AI directly" button appears to instantly re-run it in Web mode.

What is the Web search toggle in General mode, and why is it sometimes locked on?

As of W1.25, General mode has its own Web search toggle next to the model picker — on by default. With a non-search model selected, it's a normal on/off switch you control. With a search-native model selected (any Perplexity Sonar), the toggle is forced on and disabled — a lock icon and a reason line explain that model searches the web on every request whether or not you ask it to, and name DeepSeek V4 Flash as the model to pick instead if you want a General answer with no web search. This exists so the toggle never shows "off" while a search happened anyway — what it displays is always what actually ran. A General answer with web search on is a paid provider call, same as any other web search action.

What is the Thinking toggle in Ask Your Cafe?

As of W1.35, a Thinking switch sits right next to the Web search toggle in the composer — on General, Grounded, and shared-link Cafe chat alike. Some AI models can pause to reason through a question before writing their answer ("thinking"); the toggle is off by default and lets you turn that reasoning on for models that support switching it.

Why don't I see a Thinking toggle for some models?

It only appears for a model that both thinks by default and can be told to stop. Models that never think at all, and models that always think and can't be reliably told not to, show no toggle — for the second group, asking them to think anyway would leak their raw reasoning straight into the visible answer with no clean way to remove it, so Coffeescribe doesn't offer the switch rather than risk that.

Does turning Thinking on cost more or take longer?

Yes to both. With Thinking on, the model reasons before it answers — typically 30 seconds up to a couple of minutes, versus a few seconds off — and it uses more tokens, since the reasoning itself is billed. Off is the default because it's the fast, cheap choice for most questions; turn Thinking on when a question genuinely needs the extra deliberation.

What does the 🧠 "thought" marker on an answer mean?

It means that specific answer ran with Thinking on. Coffeescribe records the mode each answer actually used, not just what the toggle happened to be set to, so the marker only appears when the model genuinely reasoned first — including on answers reloaded from history. No marker means the answer ran with Thinking off, or predates this feature (W1.35) — both look exactly like every answer did before.

Does my Thinking choice apply everywhere, or just one Cafe?

Just one Cafe at a time. Coffeescribe remembers your last Thinking choice per Cafe, the same way it remembers Web search — turn it on in a research-heavy Cafe and it stays on there even while your other Cafes stay off.

Where did my old chat go? I don't see it in the chat window anymore.

It's saved, not gone — as of W1.35, look under the Chats tab in the left panel (as of W1.42, one of three tabs — Chats · Reports · Sources — at the top of that panel). Every conversation you've had in a Cafe lives there, newest first, and clicking one reopens it with every question and answer intact. If you had Q&A history before this feature shipped, it was automatically saved as one archived chat the first time you opened the Cafe after the upgrade, so nothing from before is missing either.

Does the AI remember my conversation in a Cafe chat?

Yes, but only within the one chat you're in. A follow-up question like "tell me more about point 3" can reference up to the last 20 exchanges of the chat that's currently open (W1.35 P3, up from 10), but it has no memory of a different saved chat — switching to another one (or starting fresh with New) means the AI starts that conversation with a clean slate. This is deliberate: an old conversation about one topic can't leak into a new one about something else.

Does my chat's memory survive a page reload?

Yes, as of W1.35 P3. The AI's memory used to live only in your browser tab and reset on reload; it's now rebuilt on the server from the chat's saved turns every time you ask a question, so reloading the page or reopening a saved chat later doesn't reset what it remembers.

If I add the chat I'm in to my cafe memory, does it cite itself?

No — as of W1.35 P3, a chat's own transcript is deliberately left out of what its own Cafe-sources answers retrieve, even when that chat is in your cafe memory. Otherwise the AI could end up quoting its own earlier reply back at you as if it were a separate research source. Other chats you've added stay fully citable — this only affects a chat's answers about itself.

Why did my answer disappear and start again partway through?

That's an automatic retry (W1.35 P3), not a bug. On the rare occasion a Web-mode answer starts streaming in and then the provider goes quiet, Coffeescribe discards the partial text and asks the question again rather than leaving you with a failed message. You may briefly see the in-progress answer vanish and a new one begin — you're only charged once, for the answer you actually receive.

What is cafe memory?

"Cafe memory" is the name for opting a saved chat into your Cafe's own research material — once it's in your cafe memory, a Cafe sources answer can cite that whole conversation the same way it cites a source you gathered. (Before W1.42, this was called "Use in answers"; you may still see that older name in a few places.) Every chat has a toggle (Sparkles icon) on its row in the Chats tab — off by default. Turn it on and an "in memory" pill appears on the row, and the button reads "Remove from your cafe memory"; from then on, Cafe-sources questions can pull from and cite that conversation (except the chat's own answers about itself — see the question above). Turn it off and the conversation stops being citable — the chat itself, and everything in it, is untouched either way. Cafe memory only ever affects Cafe-sources answers — it does nothing in Web mode, which never looks at your café's own material.

How do I start a new chat without losing my old conversation?

Click New in the composer. The conversation you were in is saved (archived) into the Chats section — titled from its first question — and the panel clears for a fresh chat. Nothing is deleted; New only used to mean "wipe everything," but as of W1.35 it means "save this one and start another."

What's the difference between "Delete this chat" in the panel and deleting a chat from its row?

They both delete a chat and its turns behind a confirmation — the difference is which chat. The trash icon in the panel (beside New) always deletes the conversation you're currently looking at. The delete control on a row in the Chats section deletes that specific chat, whether or not it's the one you're in. Deleting the chat you're currently in (either way) resets the panel to a clean new chat; deleting any other chat leaves you exactly where you were.

How do I attach a file to a café chat?

As of W1.35 P4, click the paperclip icon in the Ask Your Cafe composer (next to the question box) — it's there in both Web and Cafe sources mode. Choose Upload a file to pick something from your device, or From café sources (Web mode only) to pull in a source already gathered in this Cafe instead of uploading a new one. See What sources can I add to a Cafe? above.

What file types and sizes can I attach in a café chat?

One "Upload a file" option covers everything — you don't pick a type first. Documents (PDF, DOCX, TXT, MD, EPUB, .html/.htm) up to 25 MB, images up to 10 MB, audio up to 15 minutes, and video up to 3 minutes (audio/video are limited by duration, not raw file size — CoffeeScribe shrinks the file before sending it on).

Does an attached file work differently in General mode vs. Grounded mode?

Yes. In General mode, the file also stays with the chat — it's listed under that conversation in the Chats section of the rail, and its full text is available to every later question in that same chat, so follow-ups like "summarise this" work without re-attaching. In Grounded mode, the file becomes an ordinary café source and nothing more — it can be cited in a Grounded answer through the same ranked retrieval as any other in-scope source, but it isn't bound to the chat that uploaded it.

Why is Send greyed out right after I attach a file?

Because the file is still uploading or being converted (media files are transcribed/described by AI before they're usable). Sending while that's in progress would send your question without the file actually available to answer from, so Send stays disabled — with a "Your file is still uploading" note — until the attachment finishes and appears as a chip.

Why did my answer cost more after I attached a document?

A large attached document is read again on every question in that chat, not just the one where you attached it — so an answer in a chat carrying a large document typically costs roughly 2× a normal question. This is expected: it's the trade-off for the model genuinely reading the whole document each time rather than a partial excerpt that might miss what you're asking about.

What does "This chat is getting close to the model's context limit" mean?

It means the conversation (its history plus any attachments) is approaching how much text the currently-selected model can process at once. It's a warning at first — you can keep going — naming two ways out: start a new chat, or switch to a model with a larger context window. If you keep going past that point, sending is blocked outright until you start a new chat; your typed question is kept in the box, not lost.

What happens to a file I uploaded if I delete the chat it's in?

You're asked. Deleting a chat that carries files shows a choice: keep them (the default — they stay as ordinary café sources you can still search and read) or delete them too (removes the files you uploaded through that chat). Either way, a source you merely picked from café sources into the chat — rather than uploaded through it — is never deleted by a chat delete, only unlinked.

What is Venice, and what are the "DeepSeek V4 Flash 0731 Fast" models in the research picker?

Venice.ai is a second AI supplier Coffeescribe runs on, alongside OpenRouter (350+ models). As of W1.37, the research model picker (Ask Your Cafe, Auto-Write, Find the Truth, Verify, Detect Conflicts) also lists a few models served directly by Venice — most notably DeepSeek V4 Flash 0731 Fast, a faster-serving build of the same DeepSeek weights that's exclusive to Venice, plus the regular 0731 and Pro builds. They carry the same price labels and Writing/Research/Cost dots as every other research model, and go through the same rating process — picking one is an ordinary model choice. Venice models only appear in Research Cafe pickers, never in book-creation, Quick Read, or Import pickers.

Can Venice models search the web?

Yes, as of W1.39. Venice models used to have the Web search toggle forced off and disabled in General mode — that's no longer true. Pick a Venice model (e.g. DeepSeek V4 Flash 0731 Fast) in General mode and the toggle behaves like it does for any other non-Sonar model: a normal on/off switch, on by default. Venice runs its own live web search directly, not through OpenRouter's plugin, and returns its own citations rendered the same way as any other General answer's. Everything else about a Venice model works as before — writing, analysis, and Grounded mode (which is pinned to its own model and never searches anyway) are unaffected.

Why does a Venice search answer sometimes cost a different amount of tokens than another one?

Because as of W1.39, a Venice search answer is billed from what Venice actually reports charging us for that specific call, plus Coffeescribe's standard margin — not a flat per-search fee. Costs vary a little call to call, the same way any metered service does, so you may see slightly different token amounts between two otherwise-similar Venice search answers. OpenRouter search answers (Sonar, or any other OpenRouter model with search on) are unaffected — they still use the existing token-math + flat search-fee billing. If you're using your own Venice key (BYOK), you aren't charged Coffeescribe tokens for the call at all.

Can I use my own Venice API key?

Yes. Venice is one of the seven BYOK providers — add a key in Settings → API Keys and Research Ask-AI, Auto-Write, Find the Truth, Verify, and Detect Conflicts draw on your own Venice account whenever you pick a Venice model, the same fallback-if-it-fails behaviour as every other provider. See Using your own API keys (BYOK).

Why do some of my Ask-Your-Cafe references show up as amber chips instead of blue?

Because they came from the AI's own web search in General mode, not from your café research. Grounded answers cite your café sources and tray items as blue chips. When a General answer searches the web, those references render as amber chips explicitly labelled "The AI's own web sources — not your café research" — a deliberately different color and label so you're never left guessing whether a citation came from research you gathered or from the open web. Click an amber chip to open the page it points to.

Which mode is Ask Your Cafe in by default?

General ask AI — the left button — with web search on, so by default you get a web-search-grounded quick chat (W1.32A). Switch to Grounded from cafe sources any time you want answers drawn only from your own research; whichever mode you pick is remembered per Cafe (as of W1.44 — previously one flip changed every Cafe you own) and persists across reloads. One exception: highlighting text and choosing Ask AI from the popover pre-selects Grounded for that question, because "explain this thing" usually means "from my research" — unless you've already switched modes by hand in this session.

Why does my question jump to the top of the chat while the answer is being written?

That's the streaming view (W1.32A). In General mode the answer arrives live, word by word — so when you send a question, it pins to the top of the visible area and the answer writes itself in underneath, with formatting rendering as it arrives. The view deliberately doesn't auto-scroll while the text grows: your question and the start of the answer stay in view, and you scroll at your own pace if the answer runs long. When the answer finishes, it settles into a normal answer card (with its reference chips and meta line) without the view jumping.

What does the time on an answer mean (e.g. "· 27s")?

It's how long the answer took, wall-clock, from sending your question to the last word — shown on the meta line under each answer next to the model and token count. Web-searched answers usually take longer than plain ones because the search happens before the writing starts. Answers from before this was added (W1.32A) simply don't show a time.

Where do the numbered [1], [2] markers inside an answer come from?

When a General answer uses web search, the AI is asked to cite at sentence level, and Coffeescribe additionally anchors each citation to the sentence its excerpt actually supports — so a source you never explicitly asked about can still get a marker if the answer clearly draws on it. As of W1.39, those markers render as the same small clickable [N] buttons Grounded mode uses (amber, opening the source page) rather than as inline text links — so a Venice model and an OpenRouter model cite in exactly the same visual style on the same screen. Click a marker to check a specific statement without hunting through a reference list; the amber chip row below the answer is the same sources collected in one place.

Why doesn't the answer end with a "Sources" list anymore?

Because the reference chips (and the numbered [N] markers in the text) are the source list. The model still writes a trailing "Sources" section internally (it's part of how sentence-level citing is enforced), but displaying it would just duplicate the chips — so it's hidden, and as of W1.39 that's true everywhere the answer can appear: the live streaming answer, the saved answer, scrolled-back chat history, and anything you copy or save to your Tray or Notebooks. Nothing is lost: every source in that hidden list appears as a clickable chip or marker.

Why does the wait indicator say "Searching the web…"?

Because that's genuinely what's happening. With web search on in General mode, most of the wait before the first words appear is the live web search — so the indicator says so instead of a vague "thinking" message. Without web search it says "Asking the AI…", and in Grounded mode "Searching your research…".

Can I choose which AI model to use for research?

Yes — every signed-in user gets a model picker used for most AI operations in that Cafe — web search, paper analysis, Auto-Research, Auto-Write, Find the Truth, Verify, Detect Conflicts, and General Ask-AI answers. As of W1.25, it's a single picker that lives in the Ask Your Cafe composer (the input area at the bottom of the reading pane) instead of the search bar — change it once and every one of those surfaces uses the new choice. The picker pins Perplexity: Sonar at the top with a Recommended chip (it searches the web and cites what it finds, which is what General answers are for), then ~12 curated alternatives, then a collapsed advanced disclosure for the full 350+ catalogue (at your own risk). Each result card shows which model was used and how many tokens were charged. Grounded Ask-Your-Cafe answers are the one exception — they ignore this picker and always run on a fixed model (see the question above).

Why is a model recommended for research but not when I create a book?

Because they're two different lists, on purpose. Research and book writing want different things — a Cafe needs live retrieval and careful citation, while book writing needs good prose — so the models recommended for research are chosen separately from the ones recommended everywhere else. Changing one list never moves the other.

What do the Writing, Research and Cost dots actually mean?

They don't all mean the same kind of thing. Cost is exact — calculated from the model's real current price every time you see it, where 1 dot is the cheapest of the recommended research models and 5 the most expensive. Writing and Research are relative positions within that set, not absolute quality scores — the strongest of the recommended models shows 5 dots and the weakest shows 1. They answer "which of these should I pick", not "how good is this model in the world". Hover any rating to see what it means, who ranked it and when.

Why did a model's dots change when I didn't touch it?

Because Writing and Research are positions relative to the other recommended models. When a model is added to or removed from that set, everything is re-ranked — so a model that was the best of ten can become the second best of eleven without having changed at all. Cost never moves for this reason; it only changes when the real price does.

Who decides the Writing and Research scores?

The models are tested, not looked up. Each recommended model researches the same three claims — one easy, one medium, one hard — and writes each up as a short cited report. Every citation in those reports is then checked by actually fetching it, so dead links are a measured fact rather than an opinion, and the marker sees the title of every page cited so it can tell a genuinely relevant source from padding. The reports are marked blind, as "Model A, Model B", by a language model — which one is shown on hover with the date. We disclose that because a model judging other models is a real limitation and you should be able to weigh it.

Does citing more sources give a model a better score?

No — often the opposite. A model that cites twenty-five loosely related pages scores worse than one citing eight that genuinely address the question. In our own testing the model that produced the most working citations finished near the bottom, because they were mostly commercial and consumer sites rather than reviews or primary studies. Quality of sourcing is what counts, not quantity.

Why does a model show no Writing or Research dots at all?

Because it hasn't been rated yet, and it says "not yet reviewed" instead of guessing. Earlier versions estimated quality from a model's price, which produced confident-looking scores that meant nothing. Showing nothing is the honest answer.

Do the dots tell me whether a model's citations are trustworthy?

No, and that's deliberate. Citation accuracy is the weakest area of every current frontier model, and independent published measurement of it barely exists — so a dot would imply a precision nobody has. Where we've observed a particular model's citation behaviour ourselves, it's written out in plain words in that model's info panel instead.

Does web search work with all models?

Yes. Web search works with every model in the selector, not just Perplexity Sonar. The AI receives real-time web results regardless of which model you choose. Different models may interpret and present web results differently — experiment to find what works best.

Can I edit tray item text?

Yes. Click on any tray item's text to open an inline editor. You can modify the content directly in the textarea — trim irrelevant parts, fix wording, or add your own annotations. Press Ctrl+Enter (Cmd+Enter on Mac) to save your changes, or Escape to cancel. Each tray item supports up to 100,000 characters.

How do I add notes to my Serving Tray?

Click the + button in the Serving Tray header to create a blank editable note. This lets you paste in content from external sources (emails, other tools, your own writing) or write observations alongside your research. The note is saved as a regular tray item that you can reorder, edit, combine, or use in your Workspace like any other item.

How do I combine tray items?

Click the Select toggle in the Serving Tray header to enter selection mode. Checkboxes appear on each item — select two or more, then click Combine. The selected items are merged into a single item with --- separators between each original piece. An undo option appears for 5 seconds after combining. This is useful for assembling related excerpts from different sources into one consolidated reference.

Can I export a Serving Tray item?

Yes — as of W1.25, open any tray item in the reading pane and click Download in the pane's footer for PDF, HTML, Markdown, or TXT. This works for every kind of tray item: report-backed snippets, source-backed items (a whole source or a highlighted passage), saved Q&A answers, free notes you typed yourself, and combined items (see the question above). Before W1.25 most of these kinds had no export path at all — now the same four formats are available regardless of what the item is. It's free, no token cost, for every signed-in user.

What identifier types can I paste into a Cafe?

The identifier input (alongside the search bar in a Cafe) auto-detects: DOI (e.g. 10.1126/science.1099196), ISBN (e.g. 978-0-525-55947-4), PMID (e.g. 12345678), arXiv ID (e.g. 2305.12345), ISSN (e.g. 0036-8075), URL (including bare domains like www.bbc.co.uk). The submit button is disabled when the type is unrecognised. Identifier paste is available to every signed-in user.

What citation styles does CoffeeScribe support in Research Cafes?

Eight styles are available in the Cafe Settings panel (gear icon): APA 7 (default, via citation-js), APA 6, MLA 9, Chicago Notes-Bibliography, Chicago Author-Date (new in W1.2), Harvard Cite Them Right 2025, IEEE (new in W1.2), and Vancouver (new in W1.2). The style applies to all enriched source cards in that Cafe and to the bibliography of any Auto-Write Report. Change it any time in Settings — the change takes effect immediately on re-render. This is a per-Cafe setting, so different Cafes can use different styles. You can also override the style per-export on the Export Modal when you export a bibliography (e.g. report is APA 7, export as MLA 9).

Why are some IEEE / Vancouver citations rendered with the full journal title instead of the abbreviation?

IEEE and Vancouver styles ship with curated journal-abbreviation lookups — about 60 entries for IEEE (top IEEE Transactions) and ~180 for Vancouver (common medical journals). If your journal isn't in the curated list, the renderer falls back to the full title instead of inventing an abbreviation. The full NLM/MEDLINE list (~30k entries) is deferred to a follow-up so we don't bloat the main Cafe bundle. If you regularly cite a journal that's missing, let us know via feedback — we'll add it.

What is Auto-Write Research Report?

Auto-Write turns a Research Cafe into a fully-cited research output — a prose report, a brief, a comparison table, a slide deck, or a data spreadsheet — composed from sources you've already gathered (it doesn't research a fresh topic itself; that's Auto-Research's job). As of W1.44, click ✦ Deep on the chat composer and choose Auto-Write under the Write heading to open its setup in the centre column. Pick one of four modes (From Sources, Café, Tray + Sources, or From Tray) and optionally select an output format and cross-cutting options. Reports are saved as artefacts in the Cafe and are URL-addressable at /cafe/[cafeId]/artefact/[artefactId]. See Auto-Write Research Reports for the full walkthrough.

What output formats does Auto-Write support?

Auto-Write now supports five output formats — select in the dialog before generating:

All formats use the same citation pipeline — [N] markers and a References section are present in every format. Four cross-cutting options compose into any format: answer-first vs narrative structure, confidence signalling, "What we couldn't determine" section, and a word cap.

Can I change the citation style on a report after generating it?

Yes. Open the full-screen report viewer (click View Report on the report card), then use the Citation style dropdown and click Apply. The bibliography re-renders in the new style and in-text markers convert for author-date styles. This is non-destructive — the raw markers in the stored body are never rewritten, so you can switch styles freely.

Available styles: APA 7, APA 6, MLA 9, Chicago Notes-Bibliography, Chicago Author-Date, Harvard, IEEE.

Can I click [N] markers in a report?

Yes. In the full-screen report viewer, clicking any [N] in-text marker scrolls to and flashes the corresponding source in the Sources column. On shared/standalone report pages (no Sources column visible), clicking a [N] scrolls down to the bibliography entry for that source.

Bibliography entries are also clickable — click any entry to open the original source URL in a new tab.

Can I copy or download a research report?

Yes. The report viewer has Copy, Download (.md), and Download PDF buttons in the header. Copy puts the full report text (body + bibliography) on your clipboard. Download (.md) saves a .md file you can import into Notion, Obsidian, Word, or any Markdown editor. Download PDF (W54) saves a single, book-grade A4 PDF — the same typeset design as a Quick Download scribe export — for all four report types (Auto-Write, Verify, Conflict, and Find the Truth). On Find the Truth reports specifically, the PDF follows a fixed order: verdict graphic → report body → claim/truth matrix → all citations last. It's generated entirely by code (no AI model call), free for every signed-in user with no token cost, and it works on a shared report too — anyone you shared it with can download the PDF.

What's the difference between Auto-Write's four input modes?

Auto-Write has four modes — pick from the tabs in the dialog. As of W1.18 Step 2 there's no Topic tab — Auto-Write only composes from sources you've already gathered:

Every mode also has a "Choose sources" disclosure for granular per-source selection (pin a must-include source, exclude one) and run-picking — folding a whole prior Auto-Research run, Find the Truth run, or earlier Auto-Write report in as input. See What's a "run" I can fold into a report?.

The two-pass workflow (Auto-Research → curate → From Tray) still produces the highest-quality targeted reports. From Sources is the fast lane when curation isn't needed.

Where did the Topic option in Auto-Write go?

It moved — Auto-Write no longer researches a free-text topic itself (W1.18 Step 2). Open Auto-Write on a Cafe with no gathered sources yet and you'll see a small note under the tabs pointing you to Auto-Research instead: run Auto-Research on your topic first (it gathers the sources), then either come back to Auto-Write to compose a report from what it found, or tick Auto-Research's own Full Report toggle to write the report in the same run without the round-trip. The four Auto-Write tabs (From Sources / Café / Tray + Sources / From Tray) all still work exactly as before on whatever sources you already have.

What's a "run" I can fold into a report? (run-picking)

As of W1.18 Step 2, every Auto-Write mode's "Choose sources" panel lets you tick a whole prior run — a completed Auto-Research run, a Find the Truth run, or an earlier Auto-Write report — as a block of input, alongside (or instead of) hand-picking loose sources. Every source that run/report used gets folded into the candidate pool for your new report. This is useful when you've already done the gathering in a separate run and don't want to re-curate a Tray by hand. The composed report is saved as its own new run — it never edits the runs you pulled from.

What does Café mode write from?

Café mode writes from exactly the sources currently ticked into your Ask-Your-Café scope — the same checkboxes on the Sources rail that control what grounds a Grounded Q&A answer (see Ask-Your-Cafe Scope). If you've already curated your Q&A scope to the sources you trust, Café mode lets you generate a report from that exact same set without re-curating a separate Tray. Toggle a source's checkbox to add or remove it from Café mode's candidate pool; the tab is disabled if nothing is currently in scope.

Why are some of my sources not in the report?

A few possible reasons, most of them intentional:

• Topical relevance (W1.18 Step 2). Every gathered source is now tagged with a relevance score against the topic; sources that score clearly off-topic are excluded from what a report cites (they stay in your Cafe, tagged, as raw sources — nothing is deleted). This is a deliberate fix for occasional gather drift (see the note under Full Report).
• You narrowed the selection yourself. If you used the "Choose sources" panel to uncheck or exclude a source, or if Café mode's scope doesn't include it, it won't appear.
• The auto-selection algorithm trimmed a large set. A run with more gathered sources than comfortably fit in one report is narrowed to the strongest subset (quality, diversity, freshness) — pin a specific source in the "Choose sources" panel to force it in regardless.
• A source has no usable content. A source that's still URL-only with no readable text can't be meaningfully cited.

How much does Auto-Write cost?

Cost depends on the number of sources in your final selection — after any granular selection or run-picking. The Auto-Write dialog shows both the estimated cost and the active hard cap before you confirm. As of W1.18 Step 2, this scaling applies to all four modes (Topic mode's separate depth-scaled caps went away with the tab — depth-scaled research cost now lives in Auto-Research):

If a run would exceed the active cap, it is rejected before any cost is charged. Costs are deducted from your token balance regardless of tier.

What is the Research Intent Interview — can I skip it?

On a Deep Auto-Research run, before the run starts you'll see 2–5 quick multiple-choice questions about what you want the run to focus on, each with a "Something else…" free-text option. Yes, you can always skip it in one click — Skip runs the research exactly as it would without the feature, on the bare topic. If you answer, your answers compile into a short brief the planner sees before it drafts its research plan (and again when the report is written), so both stay aligned with what you asked. It only appears on Deep runs — Quick and Standard skip straight to research. See Research Intent Interview for the full walkthrough.

My Auto-Write report is taking a long time — did it fail?

Probably not. Cafes with more than 8 sources run a multi-pass pipeline: a planner, per-section writers, and a completeness critic — which can take 2–10 minutes depending on source count. The dialog stays open and:

• Shows a spinner for the first 3 minutes, polling for completion every 3 seconds.
• Switches to "Still writing…" after 3 minutes and polls every 10 seconds for up to 10 minutes total.
• Surfaces the report automatically when it finishes — no manual refresh needed.

If the run actually failed, you will see an error message in place of the spinner — not a silent hang. If you navigated away, come back to the Cafe; the report card appears in the Reports section as soon as the job completes.

Find the Truth has been on "Gathering evidence" or "Hunting counter-evidence" for minutes — is it stuck?

Almost certainly not. A truth run is a deep, two-pass investigation — it searches, downloads and reads sources for your claim, then runs a second full sweep deliberately hunting the best evidence against it — and usually takes 2–6 minutes end to end. The gathering phases are the slowest part. The progress panel tells you it's alive three ways: a one-line caption under the phase list says what the active phase is actually doing, an elapsed-time note appears after the first minute, and the "tokens spent so far" counter keeps climbing through every phase as the run spends. You can click Hide (keeps running) and carry on working — the verdict appears when it's done. A real failure shows an error message, never a silent hang.

Why does my run say "Waiting for a free slot" instead of starting?

Two separate caps govern deep background runs — Find the Truth, Auto-Research, and Research a Person — and either can make one wait. Your own cap limits how many of your runs execute at the same time; this is the common one, and you'll meet it by starting a second deep run while your first is still going. The platform cap limits how many run at once across everyone, so an overloaded moment degrades to waiting instead of slowing every run down. When either is full you'll see a "Waiting for a free slot" panel showing your position in line (e.g. "You're #3 in the queue"). It starts automatically the instant a slot frees up, nothing is charged while you wait, you can close the window and it holds its place, and a Cancel button lets you give up your spot at any time. The wait is bounded: if no slot frees up in time the run is released rather than hanging forever, nothing is charged, and you can start it again. Both caps apply whether you're on a platform key or BYOK, since they protect shared database capacity, not billing.

My run was released without starting — what does the message mean?

It names which cap held it up, because each one needs a different response from you:

In every case nothing was charged — a released run never spends tokens.

Why does my Find-the-Truth verdict cite different KINDS of sources — an encyclopedia AND a paper AND an old newspaper?

By design. A claim corroborated across independent source classes (encyclopedia + academic paper + contemporaneous newspaper + news) is far better verified than one backed by three copies of the same class saying the same thing — so the run deliberately selects and cites a varied set where the evidence allows it. On historical "did this happen?" claims it also always searches the digitised newspaper archives, so period press can appear alongside modern sources. Each class carries its own credibility weight in the probability breakdown (papers high, encyclopedias and newspapers medium, thin web pages low).

Why did my APA 7 (or Harvard) citation show the title instead of the author name?

That is correct behaviour for sources with no named author. Author-date styles (APA 7, Harvard, Chicago Author-Date, MLA 9) now fall back to a short title when a source has no author — for example (AI Safety Overview, 2024) — instead of leaking a bare [N] marker into the prose. The title fallback also applies to sources whose only "author" is a generic CMS byline like "admin", "Editor", or "Staff" — those are excluded from citation authorship entirely.

Why does the citation-style picker in Auto-Write now show an example?

Hovering any style in the picker shows a preview example — an in-text citation and a matching reference list entry in that style — so you can compare formats before committing. Nothing is generated until you click Run.

What is the "Research used" block on a report?

Every Auto-Write report now includes a Research used section at the bottom listing every source that informed the report — even if the report was generated as a Summary or Explainer with citations hidden. This is distinct from the bibliography (which shows only sources explicitly cited with [N] markers). Use it to trace which sources contributed to a report even when in-text markers aren't visible.

What is the Auto-Write sparse-source warning?

If any Auto-Write run — From Sources, Café, Tray + Sources, or From Tray — resolves to fewer than three usable sources, Auto-Write surfaces a sparse-source warning on the report. It means your current selection was too narrow — the report may lack depth. Fix it by widening your selection (add sources, tick more into Ask-Your-Café scope, or fold in another run), or go gather more with Auto-Research and re-running. The report is still saved either way; the warning is a heads-up, not a failure.

What is the Bibliography Editor?

Every Auto-Write Report has an Edit Bibliography button that opens the per-Report Bibliography Editor at /cafe/[cafeId]/artefact/[artefactId]/bibliography. It lists each citation as an edit card with a live preview, lets you edit fields (authors, title, journal, year, etc.), delete a citation, replace it with a freshly enriched source via the identifier-paste flow, or add a new citation manually. Manual citations create a cafe_sources row of type manual_entry. The editor is per-Report — different reports in the same Cafe can have different bibliographies.

What export formats does the Bibliography Editor support?

Four formats, plus a per-export citation-style override (defaults to the report's style but can be set to any of the 8 styles):

• BibTeX (.bib) — one file with all entries, Zotero/JabRef/BibDesk-import-ready
• RIS (.ris) — one file with all entries, Mendeley/EndNote/Citavi-import-ready
• PDF — one file per source, report-grade formatting (W62): the source's real title, a source-attribution line for transcript/web-page-derived sources, the full styled body (headings, bullets, real tables), and the formatted citation
• Markdown — one file per source for pasting into Obsidian, Notion, GitHub

The PDF format uses @react-pdf/renderer (lazy-loaded so it doesn't bloat the main Cafe bundle) and, as of W62, renders on the same book theme as report and book PDFs — SourceSerif4 serif body text with broad diacritic coverage. This is the per-source bibliography export (the same one behind the reading pane's Download ▾ → PDF button). For a whole-report PDF — the report body, verdict graphic, and matrix (where applicable) plus the bibliography, all as one document — use the Download PDF button in the report viewer itself instead (W54); see Can I copy or download a research report?.

What is Detect Conflicts?

Detect Conflicts is a background job that reads pairs of sources in your Cafe and decides whether they disagree — it works by understanding the meaning of what each source says (an AI model call, not a simple word-for-word text match), which is what "semantic" means here. It flags factual disagreements between sources. As of W1.44, click ✦ Deep on the chat composer, then Detect Conflicts under Interrogate, to start a run. The model reads pairs of sources and reports if they disagree on any checkable claim — "Source A says X; Source B says not-X". Results land as a Conflict Report in the Cafe's Reports tab. Use it before publishing a literature review to spot disagreements you'll need to address. See Auto-Write Research Reports → Detecting conflicts.

What does Detect Conflicts cover — Tray, Report, or whole Cafe?

The whole Cafe. Conflict Detect scans every cafe_sources row in the Cafe, not just Tray items or the sources in a specific Report. The pre-filter skips obviously unrelated pairs to keep cost in check, but the universe is "everything you've gathered in this Cafe". If you want a conflict report scoped to a Tray's curated set, copy the Tray into a fresh Cafe and run Detect Conflicts there.

How much does Detect Conflicts cost? Are there limits?

The cost is O(n²) with the source count (every pair must be compared), mitigated by a token-overlap pre-filter that skips clearly unrelated pairs. Hard cap: $0.50 per run — the job aborts if it would exceed that. Soft rate-limit: one run per Cafe per 5 minutes (a second click within the window is ignored). For a Cafe with ~10 sources you typically see a run cost a few cents and finish in 10-30 seconds; very large Cafes may need to be split.

What is the "N potential conflicts" badge?

After any Auto-Research run completes, CoffeeScribe automatically runs a free, lexical-only scan of your Cafe's sources for possible factual disagreements — no AI model call, no tokens charged. If it finds candidate overlaps, a calm amber "N potential conflicts" chip appears next to the Detect Conflicts button. It's a nudge ("these might disagree"), not a verdict — click it to open the same Detect Conflicts dialog, which still costs tokens if you choose to run the full AI check. The count persists with the run, so it's still there if you come back to the Cafe later.

Is the "N potential conflicts" badge the same as Detect Conflicts?

No — they're two tiers of the same idea. The badge is a free, automatic, lexical-only nudge that runs after every Auto-Research completion with no cost and no AI call. Detect Conflicts is the paid, AI-powered pass you trigger manually (see How much does Detect Conflicts cost? above) that actually reads pairs of sources and reports real factual disagreements. Clicking the badge opens the Detect Conflicts dialog — it does not run the paid check automatically.

What's the difference between Auto-Write Reports and Conflict Reports?

Both are saved as artefacts, but they answer different questions. Auto-Write Reports are research write-ups — prose with inline citations and a bibliography. Conflict Reports are triangulation outputs — a list of fact-level disagreements between source pairs with both citations linked. Verify cited reports are the cited write-ups automatically produced when a Verify claim loop completes. Auto-Write and Verify reports both live in the sources rail's Reports group (W1.6.1); Conflict Reports live in their own Conflicts group. Any can be added to the Tray to fold into a follow-up synthesis.

Where do I find the Reports section in a Cafe?

Auto-Write reports and Verify cited reports live together in the sources rail's Reports group (W1.6.1; rehomed from a card-based section to the rail in W1.13). Use Group or narrow with Show → Reports to find them, listed newest-first. Click a row to open the full report in the central reading pane — reports no longer expand inline in a card. Conflict Reports have their own Conflicts group alongside it.

Why does a verify report appear in my Serving Tray automatically?

When a Verify claim loop completes, the cited report is automatically added to your Serving Tray as well as saved in the Reports group. This happens for Auto-Write reports too. Click that tray item and its reading-pane view shows an "Open report" link at the top — click it to jump straight to the full-screen artefact viewer. You can delete the tray item if you don't want it there; the report in the Reports group is unaffected.

I ran Verify on a source and re-verified it later — where is the first report?

Both reports are in the sources rail's Reports group. It lists every report row, so the prior verify report stays reachable even after you run a second verify on the same source. The source's "View report →" link tracks the latest verdict, but all prior reports remain reachable in the Reports group (W1.6.1 fix for re-verify orphaning).

I saved a Q&A answer to Notebooks — where is my note?

As of W1.6.1, notes saved to Notebooks from a Cafe Q&A answer are grouped under the Cafe's name in the Notebooks left panel — not under Free Notes. Open Notebooks and look for the Cafe's name in the left panel (it appears between Free Notes and your Scribes). Notes saved before W1.6.1 are still under Free Notes; only new saves go to the Cafe group.

What is the "Saved from Q&A" group in the Sources rail?

When you click +Source on a Q&A answer, the promoted source is saved with the rag_answer source type (W1.6.1). In the Sources rail, these sources are grouped under "Saved from Q&A" — a proper named group based on the source type, not a title-prefix hack. The citations from the Q&A answer are preserved in the source content.

Can I add an Auto-Write Report or Conflict Report to my Serving Tray?

Yes — and as of W1.17, this also works for Auto-Research, Find the Truth, and People research reports (previously only Auto-Write, Verify, and Conflict could). Open any report in the reading pane and its Add to tray action (as of W1.44, in the single action row below the report body, not a header toolbar) pulls a synthesis snippet of the report into the Serving Tray. From there you can feed it into another Auto-Write run, reference it via @-mentions in the Workspace, or carry it through to a Scribe via the Workspace's Cafe Trays tab.

Can I save a whole report to my Notebooks, not just a highlighted passage?

Yes, as of W1.17. Every report type (Auto-Write, Detect Conflicts, Verify, Find the Truth, People research, and Auto-Research) has a Save to Notebooks button next to Copy / Download .md / Download PDF — click it and the entire report body is saved as a note grouped under this Cafe's name in Notebooks. This works from both the reading pane and the standalone report reader (/cafe/[cafeId]/artefact/[artefactId]), and it's separate from the older selection-based Save to notebook action on the Ask-AI popover, which only saves whatever text you highlighted. It's owner-only — you won't see the button on a report someone shared with you.

Does CoffeeScribe put a disclaimer on AI-generated reports?

Yes. As of W1.17, every research report — regardless of engine — shows a short line at the bottom: "This report was generated with AI. AI can make mistakes — please verify important claims against the cited references." It renders on every view, including shared and read-only reports, since the reminder matters most to a reader who isn't the one who ran the research.

Can Ask Your Cafe cite my Auto-Write, Verify, Find the Truth, or person reports now?

Yes, as of W1.17. Previously only Auto-Research's extraction cards were searchable in Ask Your Cafe (because they were embedded as Serving Tray items). Now every report type is embedded into the Cafe's search index, so a Grounded Q&A answer can cite any report the same way it cites a source or tray item — labelled "Report: {title}" in the numbered context. Click the citation to open the cited report in the reading pane.

Where is the "Add Source by Identifier" input now?

The identifier-paste input (DOI / ISBN / PMID / arXiv / ISSN / URL) lives behind a compact "Add by identifier" chip in the search bar's info row, inside the Add sources dialog (as of W1.42). Click the chip to open a drawer containing the identifier-paste input along with Cafe Settings and Providers. Paste your identifier there and CoffeeScribe auto-detects the type. (As of W1.25, the model picker is no longer in this row — it moved to the Ask Your Cafe composer; see Can I choose which AI model to use for research?.)

What happens when I try to Tray an empty-content source?

When you click the Tray button on a source card whose body content is empty (e.g. an OpenAlex paper with no abstract), a confirmation dialog now opens with three options:

• Scrape full article — fires an Apify URL scrape for the source's URL, populates the content, then adds it to the Tray. Costs a small URL-scrape fee (~3,000 tokens).
• Add anyway — adds the source as-is, just the metadata. Free.
• Cancel — closes the dialog.

This replaces the previous silent path where empty sources used to land in the Tray with nothing useful to read.

Can I share a single Auto-Write report without sharing my whole Cafe?

Yes. Click the Share button that appears on an expanded report card in the Cafe column, or click Share in the full-screen report page. A dialog opens where you toggle Enable sharing on and copy the share URL. The report link is completely independent of Cafe sharing — recipients can read the report without being able to access the rest of your Cafe.

Recipients must be signed in to Coffeescribe to open the link. The link does not work for anonymous visitors. Toggle sharing off at any time to immediately revoke access.

Can non-Coffeescribe users read my shared report?

Not currently. Shared report links require the viewer to be signed in to a Coffeescribe account. Anyone who follows the link without being signed in will be redirected to the sign-in page. Truly-public (anonymous-accessible) sharing is planned for a future update.

Where do my Auto-Write reports appear in the Cafe?

Generated Auto-Write reports persist as rows in the sources rail's Reports group (reach it via Group → Reports or Show → Reports), so they're re-openable without needing to remember or bookmark a URL. Click a row to read the full body + bibliography in the central reading pane — reports no longer expand inline in a card (W1.13). You can drag-select text in the pane to add it to your Serving Tray.

Why do some reports show citations and others don't?

Citation display defaults depend on the scribe type used when the report was generated:

A Show / Hide citations toggle is available on every report so you can override the default. When hidden, inline [N] markers and the bibliography section are both concealed — the prose reads cleanly. Toggle them on to verify sources.

Why do APA 7 reports now show "(Smith, 2022)" instead of "[1]"?

APA 7 style uses author-date in-text citations, not numbered superscripts. As of W1.3, reports generated with APA 7 citation style have their in-text markers post-processed to the correct format after the AI writes the draft: (Smith, 2022), (Smith & Jones, 2022), (Smith et al., 2022). Web sources without structured author/year metadata continue to render as [N] + URL in the bibliography. Other citation styles (Vancouver, IEEE, MLA, Chicago) continue to use numbered [N] markers as before.

What happens to tray items that don't have a source when I use From Tray mode?

Plain notes (tray items you wrote yourself or added without a linked source) are included in the report as observations — they're woven into the prose without citation markers, and they do not appear in the bibliography. Only source-linked tray items (snippets from a search result, academic paper, or uploaded document) are cited with [N] markers and listed in the bibliography.

The From Tray dialog now shows the count of source-linked items and notes separately so you know what will be cited and what will be included as observation only.

What happens when I paste a DOI or ISBN now?

After fetching citation metadata (Crossref, OpenAlex, PubMed, etc.), CoffeeScribe now automatically attempts to scrape the full article body from the source URL in the background. For open-access articles this retrieves readable content. For paywalled publishers (Nature, Elsevier, etc.) the scrape is typically blocked and the enrichment metadata plus abstract are kept as the source content. Either way the 200 response arrives immediately — the body scrape is best-effort and doesn't slow down the flow.

Why doesn't W1.2 ship the "Draft Scribe from Cafe" button?

The backend, dialog, and per-button tests for Cafe → Scribe (Draft Scribe) all shipped in W1.2 but the three entry-point buttons (Cafe header, Tray, Report viewer) are hidden behind a feature flag (DRAFT_SCRIBE_UI_ENABLED = false). They'll be turned on in a future update once we close the Workspace-visibility gap — that is, when the Scribe Workspace can render rich inline citations from the enriched book_citations rows that Draft Scribe creates.

What scope does the Bibliography Editor cover — per Report or per Cafe?

Per Report. The Bibliography Editor opens at /cafe/[cafeId]/artefact/[artefactId]/bibliography — one artefactId per Auto-Write Report. Different reports in the same Cafe can therefore have different bibliographies. To review the raw sources in your Cafe outside of any report, use the existing Serving Tray and source-card surfaces from W1.1.

What does the source quality star rating mean?

The 1–3 star score (⭐ to ⭐⭐⭐) is a heuristic signal, not an editorial judgment. It's computed from four factors: metadata completeness (full enrichment = more stars), source tier (academic journal = more stars), recency (older sources = fewer stars), and citation count (from OpenAlex when available). Hover the stars to see the numeric score (0–100) and a per-component breakdown. Use it as a quick filter — don't reject a source solely because it scored 1 star (a recent blog post about a niche topic may score lower than a 1980s journal article, but that doesn't mean the blog is wrong).

Note (W1.4, reworded W1.6.2): The star score has been replaced by a source-detail badge — Well-documented / Some research / Limited research / No details — as the primary at-a-glance signal on source cards. It describes how much we know about the source, not whether it's true. The numeric quality detail is still visible inside the badge tooltip. See below for details.

What does the source-detail badge mean on a source card?

The badge (showing Well-documented, Some research, Limited research, or No details) is an automatic, structural read of how much detail we've captured about a source — not a judgement of whether it's accurate. The weak end is grey, never red: "Limited research" means we've only got a little about it (often just a link), not that it's fake. It looks at: any academic identifier (a DOI/PMID/arXiv id, even in the URL, marks a peer-reviewed work as authoritative), the source tier (academic journal vs social post), whether a named author and a publication date are present, domain reputation, and how complete the metadata is.

The badge updates automatically when a source is added or enriched — no action is needed. Hover it to see the reasons behind the rating.

Does the source-detail badge mean the source has been fact-checked?

No — and that's the key thing. "Well-documented" means we have a lot of detail about the source (named author, DOI, academic tier, reputable domain). It does not mean any claim in it has been checked against other sources.

Whether a claim actually checks out is the separate Verify feature: every source card has a working "Verify claim" button that cross-references the claim against other, independent sources and returns a verdict (see the next entry). The source-detail badge is a fast structural screen, not a truth guarantee — always read the source (and Verify important claims) before citing it.

What does "Verify claim" do?

Verify claim runs a real cross-reference research loop on a source in your Research Cafe. When you click it, a dialog opens that:

1. Shows the source's main checkable claims (extracted by AI in a cheap preview pass).
2. Lets you pick which to verify — the main one is pre-checked, or select "verify all", or type your own.
3. Lets you set a token ceiling — the agent searches within that budget.

After you confirm, the dialog closes and the card shows "Verifying…". When the loop finishes, the card badge resolves to one of four verdicts: Verified, Contradicted, Mixed, or Couldn't verify. Click View evidence on the card to see the supporting and opposing sources the agent found, each saved as a real source in your Cafe.

Both verify loops are available to every signed-in user. See Verify Claim & Find Credible Source for the full walkthrough.

Does "Verify claim" now produce a full report?

Yes. As of W1.6, the Verify loop automatically produces a cited report in the Reports section of your Cafe when it completes. You do not need to click anything extra. The report has one section per verified claim — a heading (the claim), a verdict line, 1–2 paragraphs with inline [N] citations drawn from the evidence the agent found, and a References list containing only the sources actually cited in the body.

The card also keeps the inline evidence panel (click "View evidence") which shows the raw underlying/supporting/opposing sources. The two are complementary — the evidence panel is the detailed agent output; the cited report is the synthesised readable version.

See Verify Claim & Cited Report for the full walkthrough.

What does "Verify claim" produce — a verdict or a report?

Both. The verdict badge (Verified / Contradicted / Mixed / Couldn't verify) appears on the source card. The cited report is saved automatically to the Reports section. The inline evidence panel (click "View evidence") shows the underlying sources the agent checked. All three are produced by the same run at no extra charge.

What do the verify verdicts mean?

What does "Couldn't verify" mean — did it fail?

No. "Couldn't verify" is an honest answer, not a failure. It means the agent searched for evidence and could not find enough to reach a confident verdict. This can happen because:

• The claim is an opinion or subjective statement — not falsifiable.
• The claim is too niche or too recent — peer-reviewed literature doesn't yet exist.
• The claim is simply underdocumented — even a thorough search found nothing definitive.

A "Couldn't verify" card shows everything the agent checked — which sources it found and why it couldn't reach a verdict. This is useful context before deciding whether to cite the claim.

Does "Verify claim" in a Research Cafe cost tokens?

Yes. The verify loop runs web and academic searches, which costs tokens from your balance. The exact estimate is shown in the confirm dialog before you commit — no tokens are charged unless you click Verify. The cost is similar to a Quick Auto-Research run on a single focused claim. Verifying multiple claims or raising the token ceiling costs proportionally more.

Is "Verify claim" in the Cafe the same as "Verify Validity" in the Workspace?

They are different features:

• Verify claim (Research Cafes) — cross-references a claim in a research source against external evidence. The loop finds corroborating and opposing evidence from the web and academic databases and saves them as real sources in your Cafe. Available to every signed-in user.
• Verify Validity (Workspace) — fact-checks claims in your scribe content and returns verdicts with web source links. Available to every signed-in user. The results are saved to Notebooks as verification notes.

Use Verify claim when you want to know whether a source you gathered is accurate. Use Verify Validity when you want to fact-check the scribe you are writing.

What is "Find the Truth"?

Find the Truth (W1.11) is a verb in the ✦ Deep picker on every Research Cafe's chat composer (as of W1.44, under the Interrogate heading — a typed claim in the composer carries straight into it), available to every signed-in user, that turns any claim — or question — into a researched probability that it's true, plus a cited report. Unlike Verify claim, which checks a claim already sitting in a source you gathered, Find the Truth researches a claim you type from scratch: it gathers supporting evidence, separately hunts for the strongest evidence against the claim, weighs everything by source credibility, and returns a probability, a plain-language "What we found" answer, an evidence breakdown, and the strongest counter-evidence it found. Its setup opens fresh every time you return to it (unlike Auto-Research, Auto-Write, and Detect Conflicts, which keep a half-filled draft for the rest of your session) — a completed check stays reachable under "Past checks" either way. See Find the Truth & Claim Evidence Matrix for the full walkthrough.

What does the Find the Truth percentage mean? Why isn't it ever 0% or 100%?

The percentage is an evidence-weighted estimate, not a certainty — it's clamped between 5% and 95% on purpose. Even overwhelming evidence stops short of absolute certainty, and even a claim with almost no evidence found never rounds down to a flat zero. The number is computed by a fixed, deterministic formula from the weighed evidence (anchored on the model's overall reading of the sources) — it is not a raw "ask the AI what it thinks" guess.

Why did my claim only score in the mid-80s even though I know it's true?

Check the evidence breakdown for the credibility tiers behind the score. Sources are weighed by tier — peer-reviewed/authoritative sources carry the most weight, then established news, then general web, with unknown-tier sources (which currently includes most Wikipedia sources) carrying the least. If a claim's only supporting sources are Wikipedia, several of them together still can't push the probability much past the mid-80s, even for a claim that's genuinely well-established (this is a known example: "JFK was assassinated," backed only by Wikipedia sources, currently lands around 84%). "Unknown" tier means we haven't assessed this source — not that it's untrustworthy.

What sources does Find the Truth use?

Web search, academic databases (OpenAlex, arXiv, PubMed), news search, Wikipedia, and newspaper archives (W1.12a). It deliberately excludes social media, Google Scholar, and manual URL scraping for a truth run specifically — a probability verdict should only draw on sources that can responsibly move it. Those other tools remain available in the Cafe generally (search bar, Auto-Research).

Why does the verdict card sometimes say "we've weighed the sources as a whole" instead of showing supporting/against sources?

This happens when the system can't confidently label each individual source as "for" or "against" the claim — most often on claims involving negation or absolute wording ("always," "never," "everyone"). Rather than risk mislabeling a source, the card falls back to a neutral, honest framing. The headline probability is unaffected — it's still computed from the same weighed evidence, just displayed without a per-source split that couldn't be trusted.

Does Find the Truth cost tokens?

Yes. Find the Truth is available to every signed-in user and costs tokens — it runs two full research passes (evidence for the claim, then a separate counter-research pass against it) plus scoring and report writing, so it's one of the pricier research actions. A cost estimate for your chosen depth is shown before you run it; nothing is charged until you confirm. The companion Claim Evidence Matrix (/research/[cafeId]/claims) is free to view for every signed-in user — it's a read view of evidence already gathered.

What is the Claim Evidence Matrix, and is it different from Find the Truth?

The Claim Evidence Matrix (/research/[cafeId]/claims, reached via Claims under Interrogate in the ✦ Deep picker — as of W1.44) is a claim-centric table showing every claim checked in your Cafe — by Find the Truth or by Verify claim — with supporting/contradicting/silent sources grouped by credibility tier and deduped to "independent roots" (so several articles quoting one wire story count once, not several times). It's free to view for every signed-in user because it's just a view of evidence you've already gathered — there's no new research cost to look at it.

What is "independent roots" on the claim matrix?

It's a dedup count for corroborating evidence: if three news articles all trace back to the same original wire story or press release, that counts as one independent root, not three. This prevents evidence from looking more corroborated than it actually is just because a story got widely republished. A heuristic pass (matching by DOI or normalized URL) runs automatically; an optional AI "Refine independence" pass (small token cost, consent required, result persisted so re-running doesn't re-charge) can resolve ambiguous same-domain clusters more precisely.

Which model does a Find the Truth run use?

Whatever model is selected in the Cafe's model picker — as of W1.25, that picker lives in the Ask Your Cafe composer at the bottom of the reading pane, not the search bar. It's the same model that drives Auto-Research, Auto-Write, Verify, Detect Conflicts, and General Ask-AI answers in that Cafe. Pick a different model there before you click Run if you want a specific one; the run's cost is billed against whichever model actually did the work. Find the Truth no longer silently defaults to a fixed platform model regardless of your selection.

Why does my verdict say "insufficient-evidence estimate" instead of a confident label?

Confident bands (true, likely true, likely false, false) require the verdict to be backed by either 3 or more independent sources (any credibility mix) or 2 independent sources where at least 2 are medium/high credibility. When the evidence is thinner than that, the verdict is held to "leans true", "leans false", or uncertain, and the small label under the percentage switches from "evidence-weighted estimate" to "insufficient-evidence estimate" — an honest signal that the number reflects too little evidence to be confident, not that the claim is actually borderline. Re-run at a deeper research level, or gather more sources first, if you want a shot at a firmer answer.

What does it mean when a Find the Truth verdict shows "Mixed" instead of a directional label?

Mixed means the evidence genuinely points both ways and the model judged the claim can't honestly be called true or false — it's its own non-directional band, not a fifth flavor of "leans true" or "leans false." It renders amber on the probability meter, the same tone as uncertain.

Before this label existed, a genuinely mixed judgement was still forced through the same seven directional bands as everything else — so a two-sided verdict could print as "leans true" or "leans false" depending on which way the underlying number happened to fall, even while the "What we found" text described the claim as mixed. That's fixed going forward, but it's not retroactive: an older Find the Truth report keeps whatever label it showed when it ran, even if the same claim would score as "mixed" if you ran it again today.

This is a different feature from Verify claim's own "Mixed" result (see What do the verify verdicts mean? above) — that's a separate, source-level cross-reference check with its own four-way outcome (Verified / Contradicted / Mixed / Couldn't verify). Find the Truth's "Mixed" is one of its probability bands, not that feature's verdict. Both describe roughly the same idea — real evidence on both sides — which is why the two tools share the word.

Why do some of my older Find the Truth runs now show as failed?

If a truth run was still "running" when the server restarted (a deploy, a crash) partway through, the background job that was supposed to finish it no longer exists — but the row used to just sit there looking like it was still going, forever. Runs stuck like that for a couple of hours are now automatically marked failed the next time the Cafe loads its run history, with the message "The run was interrupted (server restart) and did not finish." It's not a new bug — it's the fix for a run that was already dead quietly starting to say so. Just start the check again.

---

Research a Person

What is "Research a Person"?

A verb in the ✦ Deep picker on every Research Cafe's chat composer that turns a name into a compiled, cited dossier instead of checking one claim. As of W1.44, reach it from ✦ Deep under the Gather heading — it goes out and brings sources back, the same job as Auto-Research with a person as the subject, so it does not sit next to Find the Truth (which checks material you already have, under Interrogate). Type a name, confirm which real person you mean from a free candidate list, then run a paid, footprint-gated research sweep that produces an era-by-era life story, quoted social posts with real engagement numbers, and a "Who They Are" synthesis. Unlike Auto-Research and Auto-Write, its setup opens fresh every time — it doesn't keep a half-filled draft between visits. See Research a Person for the full walkthrough.

Why can't I research my neighbour (or anyone without a public profile)?

You can — but CoffeeScribe treats a private/unconfirmed person very differently from a confirmed public figure, on purpose. If your Stage-1 search doesn't turn up a matching public figure (or you pick "None of these — use my details only"), the run stays clean-core only: public web, news, and Wikipedia sources based on the details you typed. It never unlocks social-platform research for someone who isn't a confirmed public figure, and the report opens with an honest limited-coverage note explaining why. This is a hard guardrail, not a bug — see the responsible-use notice shown on every run: "CoffeeScribe never collects contact details or addresses, and limits research on people without a public profile to basic public sources. Don't use this to monitor, harass, or profile private individuals."

Why was my custom focus refused?

A custom focus (the "Anything specific to focus on?" input on the confirm step) is meant to steer the research toward a topic — "his death," "her early career." It's never treated as an instruction to the AI. If a focus asks for something location- or contact-seeking — an address, "where he lives," a daily schedule — it's declined with a polite message before it ever runs, rather than quietly ignored or (worse) actually attempted. This is the same no-contact-info hard rule that applies to the rest of the pipeline, just enforced earlier, before any spend.

Why is a dead person's social account labelled "estate-run"?

For a deceased subject, CoffeeScribe knows their still-active accounts are no longer being posted to by the person themselves — they're typically run by an estate, family, or fan/management team after death. The report labels those accounts estate-run/memorial in the Death & Aftermath section and never quotes recent posts as if they were the person's own words. This also means you won't see "posts today" framing for someone who has passed away.

What does the cost estimate on the confirm step actually cover?

It's a live quote — tokens plus any Apify (social-platform) cents — for exactly what's currently ticked: the selected depth, the lane checkboxes you've left ticked, and any custom focuses you've added. Untick a lane, remove a focus, or drop to a shallower depth and the estimate drops immediately, before you commit to anything. Nothing is charged until you click Start.

Why did my report get an "Unverified identity" caution note?

This happens on a details-only run (you picked "None of these" or nobody matched in Stage 1) where the research still turns up someone with a similar name. CoffeeScribe will never quietly assume that person IS your subject — the report frames it as "possibly related, unverified" unless what you typed substantively matches. If you do mean that specific person, it's worth re-running Stage 1 with more identifying detail so the candidate picker can find and confirm them properly, unlocking the full (not clean-core-only) sweep.

Can I add a social handle CoffeeScribe's lookup missed?

Yes. The confirm step's verified-accounts card lets you add your own handles for any platform CoffeeScribe's automatic lookup didn't find or verify — useful if you know an account that doesn't cleanly link back from Wikidata.

How long does a Research a Person run take?

Roughly 10–40 minutes, depending on depth and how many lanes/platforms you've left ticked. A "Leave no stone unturned" run across every lane on a heavily-documented public figure can take 30–60 minutes. It runs entirely in the background — close the tab, refresh, or come back later; progress reattaches automatically. If the platform-wide run cap is full when you click Start, your run may briefly show a "Waiting for a free slot" panel first — see Why does my run say "Waiting for a free slot" instead of starting?.

Does deleting a "People research" run also delete anything from my Serving Tray?

No. Deleting a person-research run (trash icon on its rail row) removes that run's report and its sources from the Cafe's sources rail, exactly like deleting any other run — but anything you've already pulled into your Serving Tray is a separate, deliberately-saved copy and is never touched.

What is "Earlier research" in the Sources rail?

Sources gathered by the older AI web-search path (before W1.4) — where the AI combined multiple results into a single synthesized response blob — are grouped under an "Earlier research" group in the sources rail (or reached via Show → Earlier research). They are preserved exactly as gathered and rendered read-only.

"Earlier research" rows do not show a source-detail badge, Verify button, or full reading-pane detail beyond the raw response (they are multi-result blobs the per-document reading view can't correctly represent). The "Well-researched only" filter hides them. Your research data is never deleted — if you have older cafes, their content remains accessible in this group. As of W1.13, "Earlier research" rows do get a scope checkbox — even though their content is read-only, they can still ground Q&A answers, so you need full control over whether they're included in scope.

What does the "Well-researched only" filter do?

The "Well-researched only" toggle in the sources rail hides sources with a Limited research or No details badge, leaving only Well-documented and Some research sources visible. It also hides "Earlier research" legacy entries. Use it when you want to see only the well-detailed sources you'd feel comfortable citing.

The toggle is visual only — it does not delete anything. Toggle it off to see all sources again.

What does the grouping knob in the Sources rail do?

The Group control switches how sources are grouped in the rail. As of W1.13 there are five options:

• Question (default when your research used fan-out sub-questions) — by the sub-question that produced each source
• Type (default otherwise) — by source type (Web, Academic, YouTube, Reddit, …)
• Used — splits into "Used" (cited in a report or added to the Tray) vs "Unused"
• Detail — groups by Well-documented / Some research / Limited research / No details
• Verify — groups by verification status
• None — a flat list with no grouping

Groups render as collapsible headers in the rail (large groups collapse by default so the rail stays scannable at 1,000+ rows). A separate Show dropdown narrows the rail to exactly one category at a time — a single type, Reports, Conflicts, or Earlier research — the same at-a-glance isolation the old pill row gave you. Both controls are purely view preferences — they change nothing about your data — and your choices are remembered per Cafe.

What is the "Used in reports" section inside a source?

Open any source in the reading pane — a "Used in reports" section appears if that source was cited in one of your Auto-Write reports. It shows the exact passage(s) the report extracted from the source — the sentence or excerpt the AI used to support a claim — tagged with which report used it.

This means a source your report cited is never a bare reference: you can always see what the report took from it, and verify it against the original. Multiple passages appear if the source was cited for more than one claim.

What is the Metadata Health badge on source cards?

The coloured pill (Full / Partial / URL only) shows how complete the enrichment data is for a source. Full (green) means all structured fields are present (author, year, journal, volume, issue, pages, DOI). Partial (amber) means some fields are missing. URL only (grey) means the enrichment returned only a bare URL — no structured metadata. The badge helps you quickly spot sources worth citing vs. sources that need manual verification.

Why do I need to consent to each Apify provider before using it?

Each Apify Actor scrapes a different platform — LinkedIn, Instagram, TikTok, Twitter, Google Scholar, Reddit, or YouTube. Each platform has its own terms of service regarding automated access. The consent modal shows a plain-language summary of the relevant risks and the approximate per-result token cost before you enable an Actor. Consent is recorded per Cafe per Actor — you only see the modal once per Actor per Cafe. You can revoke consent (and disable the Actor) at any time by toggling it off in the Provider Picker.

What is the account-level "Social Sources Terms" screen, and how is it different from the per-Cafe consent above?

As of W1.41, there's a second, wider layer above the per-Cafe/per-Actor consent described above. Social platform sources — Instagram, Facebook, TikTok, YouTube, X, LinkedIn, Reddit, Google Scholar, and smaller platforms like Telegram, Snapchat, and Truth Social — are off for your whole account by default, checked before any per-Cafe consent even matters. The first time you start research that could reach one of them (Auto-Research, Research a Person, or a manual platform search), CoffeeScribe shows a scrollable Social Sources Terms panel with three ways to leave it:

• Accept — ticks the box and clicks Accept: your acceptance (with the terms version and date) is recorded, and social platforms become available everywhere, subject to the usual per-Cafe consent.
• "Not now" — declines for this run only: your research still runs and finishes normally, just with social platforms left out. The run record names exactly which platforms were skipped, so nothing disappears silently.
• Closing the dialog (Escape, the X, or clicking outside it) — backs out of the whole action. Nothing starts, nothing is recorded, and no tokens are spent — this is different from "Not now," which does proceed.

Change your acceptance any time in Settings → Data & Privacy → Social platform sources — the switch there shows which terms version you accepted and when, and turning it back on if you'd turned it off re-opens the same terms screen if you'd never accepted (there's no per-tier or admin exception; this applies to every account the same way).

Does CoffeeScribe scrape LinkedIn?

Yes, when you explicitly enable the LinkedIn Actor in the Cafe Provider Picker and consent to its terms. The Actor is provided by a third-party Apify developer (apimaestro/LinkedIn Profile Detail) and runs on Apify's infrastructure. LinkedIn's terms of service prohibit scraping; Apify absorbs that risk on their side. Your queries are not attributed to your personal LinkedIn account. This Actor is available to every signed-in user and requires two things: your account-level acceptance of the Social Sources Terms (shown the first time research could use a social platform, and revocable in Settings → Data & Privacy → Social platform sources) and the per-Cafe consent. You can disable it at any time.

What's the maximum references per search?

Up to 25 per search, for every signed-in user — the same ceiling regardless of account. Set it in Cafe Settings; leave it on Auto (default, resolves to 7) if you don't need a fixed count.

How do citations work in research results?

When you run a web search, the AI cites its sources as clickable links within the text. A numbered Citations section appears at the bottom of each result with all referenced URLs. You can click any citation to open the original source.

How do I know where a tray item came from?

Each tray item shows a small source provenance label — for example, "Web: coffee brewing methods", "YouTube: video title", or "Academic: paper title". Items you created manually show "Manual note". These labels appear in the Serving Tray, the Workspace Cafe Trays tab, and the import modal. When you import tray items into a Scribe section, the source attribution includes the original query and URL.

Do Auto-Research's reports show which sources they were synthesised from?

Yes. Auto-Research's synthesised cards (Key Findings, Summary, Best Quotes, etc.) show a "Research used" list of every source that fed into them — click any listed source to open it in the reading pane, so you can trace every synthesised claim back to the source that supported it. As of W1.17, these cards are full reports, not Serving Tray items — as of W1.6 they showed the same source attribution as inline chips on a tray item; the underlying capability is unchanged, only where you find the cards changed (they now live under the Auto-Research run in the Sources rail, with the full report action set — see the next question).

Do Auto-Research reports get the same actions as other reports (Share, Add to Tray, PDF, etc.)?

Yes, as of W1.17. Auto-Research's extraction cards used to be the one exception in the whole app — they were stored as Serving Tray items, so opening one only gave you Copy/Edit, and "Add to Tray" was literally impossible (they already were a tray item). Now they're stored the same way every other report is, so opening one gives you the full set: Open full, Share, Add to Tray, Delete, Copy, Download .md, Download PDF, and the citation toggle — identical to Auto-Write, Verify, Conflict, Find the Truth, and People research reports. This applies to your existing back-catalogue of Auto-Research reports too, not just new runs.

What is "Ask Your Cafe" — is it still a chat bubble?

Not anymore. As of W1.13, Ask Your Cafe is the home view of the central reading pane — not a floating chat bubble in the bottom-right corner. Open a Cafe and, whenever nothing else is open in the pane (no source, tray item, or report selected), you're looking at the chat, ready for a question, with an overview strip (source count, in-scope count, conflicts badge, latest-report link) above it. Click any source, tray item, or report and the chat tucks behind it — click "← Ask your Cafe" in the pane to come back. On mobile, an "Ask" button opens the same chat. The chat opens in General ask AI mode by default (W1.32A) — a web-search-grounded quick chat whose answers stream in live; switch to Grounded from cafe sources and the AI answers using only your Cafe's in-scope material, citing sources with numbered references like [1], [2] — click any citation to open the original source in the pane. As of W1.25, Grounded answers run on a fixed model (DeepSeek V4 Flash), separate from the model you choose for web search and General answers — see What's the difference between Grounded and General?.

How do I make Ask-Your-Cafe answer only from certain sources?

Use the checkboxes on the sources rail (left side of the Cafe) — every source row (including legacy "Earlier research" rows) has a checkbox controlling whether it's included in Grounded Q&A answers. The rail header shows "N of M in scope," and each group has a toggle-all control for bulk scoping. Unlike the old session-only funnel filter, this choice is saved — it persists across reloads and sessions, not just for the current chat. Explicit "ask about THIS source" requests from the Ask-AI popover still override the checkboxes for that one question.

Why did Q&A answer when I unchecked all my sources?

Because your Serving Tray is never scope-gated — Ask Your Cafe treats tray items as already-curated and always available, regardless of the sources-rail checkboxes. If you uncheck every source but still have items in your Serving Tray, Q&A answers from the tray with an honest "Answering from 0 of N sources" banner on the reply, so you always know exactly what grounded the answer. Only when there's truly nothing in scope anywhere (no checked sources AND an empty tray) do you get the "all sources are toggled off" message instead of an answer.

Why does Ask Your Cafe show a spinner?

Your sources are being processed for search. In plain terms: when you add new research to a Cafe, CoffeeScribe breaks the text into small passages ("chunks") and builds a searchable index of them ("embedding" them) so that Cafe sources Q&A can quickly find the right passage for your question later — this whole technique is sometimes called RAG (Retrieval-Augmented Generation). You don't need to understand the mechanics, just that this prep step happens once per source and takes a few seconds. The chat is available immediately so you know Q&A is coming, and the spinner disappears once processing is complete.

Is my research data sent to external services for Q&A?

No. Your research text is turned into that searchable index locally on the server (using a model called "nomic-embed" — an in-house step, not an external service) — no research content is sent to any external embedding API. The only external call is when the AI generates an answer to your question — and it only receives the relevant passages, not your entire Cafe.

How do citations work in Q&A answers?

When the AI answers a question, it cites passages as numbered references like [1], [2]. These appear as clickable badges in the answer. Click one to scroll to the original source in your Cafe. When you copy an answer, the full reference list is automatically appended so the citations remain meaningful outside the chat.

How do I start over or clear my Q&A history?

There's no single "Clear Chat" button that wipes everything — that changed with W1.35. Click New in the composer to start a fresh conversation; your old one is saved (not deleted) under the Chats tab, titled from its first question. If you actually want a conversation gone, delete it: the trash icon beside New deletes the one you're currently in, and the delete control on any chat's row in the Chats tab deletes that specific chat — both ask for confirmation first and cannot be undone.

What are follow-up questions, and why am I charged for something I didn't type?

As of W1.42, up to six clickable follow-up-question cards appear under every answer — click one and it's asked as your next question. Generating them is a small AI call that fires automatically the moment your answer finishes, so it costs a few tokens per answer even if you never click a card. It's cheap by design: the call always runs on a fast, low-cost model chosen by Coffeescribe — never the model you picked for your chat — so the extra charge is a small fraction of the answer's own cost, and it's BYOK-aware like any other call. Don't want it? Open the model chip in the composer and switch off "Suggest follow-ups" (on by default) — no more cards, no more charge. You can also click "hide" to dismiss just one answer's cards without turning the feature off.

How much does Q&A cost?

Each question uses a small portion of your token allowance (varies by answer length). There's also a small one-time embedding cost when research sources are first added (chunks roughly every few thousand characters) — see Why does Ask Your Cafe show a spinner? above for what that means in plain terms. Embedding happens automatically in the background, and the per-action estimate is always shown before you confirm. As of W1.42, a Cafe sources answer with follow-up suggestions turned on also carries a small extra charge for those suggestions — see the question above.

Can I save Q&A answers?

Yes. Each answer has action buttons: Copy (with references appended), +Tray (save to Serving Tray — titled "Cafe chat: {your question}", citations preserved and clickable), +Source (save as a new research source — saved as a "rag_answer" source type and grouped under "Saved from Q&A" in the Sources tab), +Notebook (save to Notebooks as a cafe note, grouped under this Cafe's title in the Notebooks left panel — not under Free Notes), and Delete (remove single entry).

Do citations survive when I save a Q&A answer to the Tray or Notebooks?

Yes. All three promote actions preserve the source references from the answer:

• Save to Tray — the tray item shows clickable [N] chips. Click one to open the cited source in the reading pane.
• Save as source — the promoted source is saved as a rag_answer source type (W1.6.1) and grouped under "Saved from Q&A" in the Sources rail.
• Save to Notebooks — the note is saved as a cafe note (W1.6.1) and grouped under this Cafe's title in the Notebooks left panel. It is separate from Free Notes. Open Notebooks and look for the Cafe's name in the panel.

If you saved Q&A answers before W1.6, those older notes still exist but display without clickable citations (the references were stored as text rather than source IDs). If you saved Q&A answers to Notebooks before W1.6.1, those notes appear under Free Notes (the old path); new saves go under the Cafe's title.

Can I share a Cafe with someone?

Yes. Open a Cafe, click Share in the view bar (top of the workspace), and toggle sharing on. A read-only link is generated that you can copy and send to any Coffeescribe user. Viewers can browse your sources, tray items, and ask Q&A questions — but they cannot search, add sources, or modify anything. Their Q&A uses their own tokens and is not saved to your Cafe.

Who pays for Q&A on a shared Cafe?

The viewer pays. When someone asks a Q&A question on your shared Cafe, the tokens are deducted from their account, not yours. Their Q&A history is session-only and disappears when they close the page. As of W1.35, this includes the Thinking toggle — if a viewer turns it on for a model that supports it, the extra reasoning tokens still come out of their balance, not yours.

How do I stop sharing a Cafe?

Open the Share dialog and toggle sharing off. The link immediately stops working. You can also click "Regenerate link" to create a new URL — the old one is permanently invalidated.

Is shared Cafe content moderated?

You are responsible for the content in your shared Cafes. Coffeescribe does not review shared research content. If a viewer encounters inappropriate material, they can report it via the Report Content button on the public scribe — a modal opens with a pre-filled email template, and the viewer sends it to support@coffeescribe.ai with subject REPORT CONTENT (all caps). See the Content Guidelines for the full reporting steps. AI-generated research may also contain inaccuracies — see our AI Content Policy for details.

Can I rename research sources?

Yes. Double-click any research source title to edit it inline. Press Enter or click away to save, or Escape to cancel. This helps you organise sources when the auto-generated titles aren't descriptive enough.

Can I paste a list of links into a Research Cafe?

Yes. As of May 2026 (Epic 11.1), the Cafe search bar auto-detects when you paste two or more sources and opens a batch import modal automatically — no separate button. Paste a mixed blob of YouTube URLs, article links, free-text web queries, and arxiv: / pubmed: / openalex: prefixed academic queries (newline-, comma-, or whitespace-separated) and each item is dispatched to the correct backend in parallel. One bad URL doesn't stop the rest — failures are isolated per row, retried up to 3 times on transient errors (5xx, 429, timeout, Apify), and surfaced with the underlying error message so you can hit Retry on just the failed ones. Every signed-in user can batch up to 50 items.

How much does batch import cost in a Research Cafe?

The batch modal uses the same per-item fees as the single-input search bar — there is no separate batch surcharge. Each item is billed individually on success: ~8,000-10,000 tokens per AI Web Search, 2,000 tokens per academic search (OpenAlex / arXiv / PubMed), 5,000 tokens per YouTube transcript, 3,000 tokens per URL scrape. The modal shows a pre-flight total (sum of every parsed item) before you click Run all, and the post-run summary banner shows exactly how many tokens were charged. Failed items don't bill.

Why does the multi-source modal open when I just pasted a single URL?

The modal only opens when the parser detects two or more sources. Single-URL pastes flow through the normal search bar. If it pops unexpectedly, your pasted text probably had stray whitespace, line breaks, or trailing characters that the parser split into multiple tokens — close the modal, clean the text, and paste again. The modal is non-destructive: nothing runs until you click Run all.

What's the difference between batch import and Auto-Research?

Batch import takes the sources themselves — you already have the links and queries — and ingests them into the Cafe in parallel. It's a fast lane for a curated reading list. Auto-Research takes a topic and lets an AI agent autonomously plan, search, and synthesize across web/academic/YouTube/scrape/wikipedia/news/social tools. Use batch when you know what you want; use Auto-Research when you want the AI to figure it out. They're complementary and you can mix them in the same Cafe.

What is Auto-Research?

Auto-Research is a steerable AI research agent that searches a topic for you across a wide range of tools — web, academic databases, YouTube, URL scraping, Wikipedia, news search, and optionally Reddit, Google Scholar, X, Instagram, LinkedIn, and TikTok. You control what it extracts, which tools it uses, how thoroughly it researches (depth), and how wide a candidate pool each search casts (results per search). The agent creates a plan, executes it, adds sources to your Cafe's Sources rail, and synthesizes curated findings into your Serving Tray.

What is Quick mode (Auto-Research's fast answer)?

As of W1.32B, Auto-Research has a true fast mode, and it's the default when you open the setup: Quick research (as of W1.44 — previously labelled "Quick answer"; the wording changed, the behaviour didn't). Ask it anything and it plans its own research in seconds — it picks the best specialist gatherers for your question (social platforms, academic papers, news archives, the live web) from a wide shelf of tools, runs them all at once, and hands you a fully cited report a few minutes later, with every source saved into your café. It runs on genuinely different, faster machinery from Solid research and Leave no stone unturned (the renamed Standard/Deep) — see the next question for how they differ. Full walkthrough: Research Mode → Quick Mode.

What's the difference between Quick mode and Solid/Deep research?

Quick plans once (a single fast call), fires up to 7 specialist gatherers plus a live web search all in parallel, and always finishes with a cited report — typically in a couple of minutes. It has a simplified setup: no Advanced panel, no manual tool picking, no fan-out or prompt editing. Solid research and Leave no stone unturned (previously labelled Standard and Deep) run the original, more configurable agent — a step-by-step loop with a full Advanced panel (manual tools, results-per-search, fan-out, editable prompts, research skills), and Leave-no-stone-unturned can additionally reflect and replan across extra rounds. Reach for Quick when you want a fast, good-enough answer; reach for Solid/Deep when you want to steer the run yourself or need genuinely deep, many-source coverage.

Can I skip the report in Quick mode and just get sources?

No — a Quick run always writes a full cited report; there's no "sources only" option for Quick. That trade only exists for Solid/Deep, where Advanced → "Research only (no report)" turns the report off and leaves you with just the gathered sources and extraction cards.

What are the new Auto-Research tools in W1.5b?

Auto-Research now supports ten tools total. The four always-free tools (Web, Academic, YouTube, URL Scrape) are joined by Wikipedia (free, no consent needed — MediaWiki REST, good for orientation and definitional queries) and News search (Apify actor, returns recent article titles and outlets). Six consent-gated tools (every signed-in user, token-billed) are now wired and returning real results: Reddit (community discussions), Scholar (Google Scholar via Apify), X (recent posts), Instagram, LinkedIn, and TikTok.

Can I search old newspapers?

Yes. Newspaper archives (W1.12a) is an always-free toggle in Auto-Research's tool list that searches digitised historic newspapers and periodicals — good for "how was this reported at the time," obituaries, old local history, and claims that predate the internet. It's the archival counterpart to the present-day News tool. Coverage comes from Internet Archive and Europeana; Chronicling America (Library of Congress) is wired in but currently Cloudflare-blocked from our servers, so results come from the other two providers for now. See Research Mode → Newspaper Archives for the full walkthrough.

Why are some newspaper results snippet-only?

Copyright. Newspaper archives never exclude a result, but how much text is stored depends on its rights status: public-domain or openly-licensed articles are stored in full, in-copyright items the archive lets us preview are stored as a snippet + link, and in-copyright items with no preview are stored as a title/publication/date + link only. Every tier is a real, citable, honestly-labeled source — you can always open the original link to read the rest.

Which countries/archives does Newspaper archives search?

Internet Archive (broad, largely US-centric digitised newspaper and magazine collections) and Europeana (European national and regional newspaper collections, aggregated across many countries' digitisation projects). Chronicling America (US, via the Library of Congress) is implemented but currently blocked server-side by Cloudflare — it will resume automatically once that's resolved, no action needed from you. Trove (Australia) and a few other national archives are known v1 gaps, tracked as a follow-up.

Does a newspaper source count as strong evidence in Find the Truth?

It depends on what you're asking. A newspaper article is strong evidence for whether something happened or was reported at the time — but weak evidence on its own for whether an underlying claim is actually true (a period newspaper repeating a contemporary rumor doesn't make the rumor true). Find the Truth's scoring applies this distinction automatically per claim, and snippet-only/reference-only newspaper sources are additionally capped so they can't single-handedly push a verdict.

How much does Auto-Research cost?

It depends on depth, tools, and results-per-search. Quick (as of W1.32B, the fast engine — see What is Quick mode? above) is typically modest: one planning call, up to 7 gatherer lanes plus a web search run in parallel, and its always-on report. Solid research uses a larger portion of your allowance, and Leave no stone unturned can vary widely depending on how many steps the agent decides to take. Every depth, Quick included, shows a token budget field before starting — this is a hard ceiling, not a target — and the pre-flight estimate in the setup shows the projected cost before you commit.

Will a Deep run burn my whole token budget?

No. "Deep" tells the agent to research as thoroughly as it judges useful — but the token budget is a ceiling, not a target. The agent uses its own judgment and typically stops well before the budget is exhausted, once it believes it has gathered enough. The budget just ensures the run cannot accidentally exceed what you're comfortable spending. A Deep run on a focused topic may use far fewer tokens than the budget ceiling.

What's the difference between research depth and results per search?

They control two different things, and both are Solid/Deep-only as of W1.32B — Quick has no Advanced panel and no results-per-search control at all (see What is Quick mode? above):

• Depth (Solid research / Leave no stone unturned, previously called Standard / Deep) — how hard and how many cycles the agent runs. "Leave no stone unturned" = most thorough; the agent stops when it judges it has enough.
• Results per search (5 / 10 / 20, under Advanced) — the candidate pool each individual search fetches. Higher top-k means the agent sees more candidates per step, which is more expensive but gives it more to choose from.

For example: Solid research + 20 results-per-search runs fewer cycles but each cycle considers a wider candidate pool. Leave-no-stone-unturned + 5 results-per-search runs many cycles but each looks at fewer candidates.

Why do Reddit, Google Scholar, X, Instagram, LinkedIn, and TikTok need consent?

These tools use Apify actors to access content on third-party platforms that have their own terms of service regarding automated access. Before enabling any of these tools in Auto-Research, a consent modal explains what data is sent, the approximate per-result cost, and links to the platform's terms. Consent is recorded per Cafe per tool — you only see the modal once per tool per Cafe. The executor also enforces consent server-side as a backstop. These tools are available to every signed-in user — consent and token billing still apply.

Why is Google Scholar content sometimes short?

Scholar results are returned as snippets by the actor's design — the full PDF is not retrieved automatically. The snippet is enough for the agent to assess relevance and cite the source. If you need the full text, use the "Fetch full article" option on the source card, or paste the DOI into the identifier input to trigger a full enrichment.

What are the extraction presets, and what does "picked for you" mean?

Under "What to EXTRACT," you choose which synthesized cards the run produces at the end — eleven presets as of W1.32B (grown from five): Key Findings, Summary, Plain-English Explainer, Best Quotes, Latest Developments, Contradictions & Gaps, Timeline, Statistics & Data, Expert Opinions, Pros & Cons, and Action Items. Each preset now has its own distinct instruction and length target, so ticking several gets you genuinely different cards, not near-duplicates with different labels. You don't have to pick anything — leave the pills alone and a sensible default set is picked for you automatically. Want something none of the eleven cover? Describe it in the box just below the pills.

What is the "Go deeper?" offer after a Quick run?

When a Quick run finishes, a panel with twelve next moves appears — built from your run's own topic and the sources it actually found, not generic suggestions. Pick one and it opens a new Auto-Research run pre-filled at Solid or Deep depth with the topic already set; nothing launches and nothing is charged just from clicking an option — you still hit Start to actually run it. It's generated once, right when the Quick run finishes; if the run found nothing worth building suggestions from, no panel appears at all and nothing is charged for the attempt.

What is fan-out in Auto-Research?

Fan-out (under the Advanced panel) controls whether the agent decomposes your topic into parallel sub-questions or researches it as a single linear sweep. The control is a tri-state:

Deep selects Auto by default — because a thorough run naturally benefits from letting the agent decide. Quick and Standard default to Off. You can change this at any time in Advanced, and your explicit choice is remembered even if you switch depth.

What is the "Auto" fan-out setting and when should I use it?

Auto means the planner LLM decides at runtime whether your topic benefits from decomposition into sub-questions. If it does, it creates parallel research groups; if not, it runs a single linear sweep. Use Auto when you're not sure whether your topic is multi-faceted — the agent's judgment is usually good. Use On when you definitely want sub-question decomposition (e.g. a literature review across multiple themes). Use Off for simple, focused topics where a linear sweep is sufficient.

Why does the cost estimate say "Rough estimate" when I use Auto fan-out?

When fan-out is set to Auto, the planner decides at runtime how many sub-questions (if any) to create. Because the shape of the run isn't known until the planner executes, the pre-flight estimate is a best-guess rather than a deterministic number. The warning "Rough estimate — the agent decides how deep to go, up to your cap of X tokens" appears near the cost estimate. The executor's hard ceiling still applies: billing charges only actual tokens used, so the run can never exceed your budget even if the estimate is off.

Why did my Deep Auto-Research run show a second "Research plan ready" step?

Deep runs can now reflect and replan. After the first research pass, the agent checks whether it left any real gaps and, if so, runs up to two more focused rounds chasing the specific unanswered facets — you'll see a second (or third) "Research plan ready" boundary in the live progress log. This is the agent digging deeper, not a glitch. If the first pass already answered the topic well, it stops after one round — that's the correct, expected outcome for a well-covered topic, not a shortfall. Reflect-and-replan only applies to Deep; Quick and Standard never do it. It stays inside the same token budget you set and is hard-capped at two extra rounds.

Does reflect-and-replan cost extra tokens?

No separate charge — the extra rounds run inside the same token budget ceiling you set for the run, they don't get a fresh budget. The budget is still a hard ceiling: if it's spent after round 1, no further round runs regardless of whether the agent would have liked to keep digging.

Can I turn off reflect-and-replan?

Not from the Auto-Research setup yet — there's no toggle in the UI today. (It exists behind the scenes and will get a control in a future update.) Every Deep run currently reflects and replans by default.

What are Research Skills in Auto-Research?

Skills are toggleable preset research playbooks under the Advanced panel. Three are available:

• Systematic Literature Review — structured, comprehensive, tracks methodology
• Fact-Check Claims — actively seeks disconfirming evidence before accepting a claim
• Find Opposing Views — seeks counter-arguments and alternative perspectives

Each skill injects additional instructions into the agent's planner and synthesis. They can be combined. You can also edit the text of any skill's instructions and reset to the default.

Can I edit the agent's prompt?

Yes. Under Advanced — prompt, skills & loop, five structured sections (Role, Method, Evidence & Skepticism, Source Hierarchy, Output Contract) are pre-filled with sensible defaults. Edit any section to change the agent's research strategy, then click Reset to restore the default. The machine-readable JSON plan contract is always appended by the system after your text and cannot be overridden — your edits change strategy, not the agent's ability to produce a parseable plan.

Does my Auto-Research keep running if I close the tab?

Yes. Auto-Research runs as a server-side background job — it keeps going even if you close the tab, refresh the page, or navigate away. Come back to the Cafe and your sources will be there when the run finishes. If a run is still in progress when you reload, the live progress panel reattaches automatically.

Will my reports (Key Findings, Best Quotes, etc.) land if I close the tab mid-run?

Yes, as long as the run completes. Sources are saved to your Cafe as each step runs; reports are added once the full run finishes (as of W1.17, these are real reports under the Auto-Research run in the Sources rail, not Serving Tray items — see Do Auto-Research reports get the same actions as other reports? above). If you cancel mid-run (by clicking Stop), sources gathered before you stopped are kept, but reports are not produced for a cancelled run — the synthesis step only runs on completion.

Can I stop Auto-Research mid-run?

Yes. Click the Stop button at any time. The server-side job receives the cancel signal and stops at the next step boundary. Sources already found are kept in your Cafe. Reports (Key Findings, Best Quotes, etc.) are not produced on a cancelled run — they are only synthesized when the run completes normally.

What happens if an Auto-Research step gets stuck?

Each step has a 90-second timeout. If a tool step (for example, a slow Apify actor cold-start) does not complete within 90 seconds, it is automatically skipped and surfaced as a skipped step in the progress log. The run continues with the remaining steps — you will never see an infinite spinner. Cold-starting Apify actors (Reddit, Scholar, social tools) can sometimes take 20-40 seconds on the first call; this is normal and within the timeout window.

If your run shows a "Waiting for a free slot" panel instead of starting at all, that's a different thing — a cap on how many deep runs (Auto-Research, Find the Truth, Research a Person) can execute at once, either yours or the platform's. See Why does my run say "Waiting for a free slot" instead of starting? above.

Where did the robot-icon items in my Serving Tray go?

As of W1.17, Auto-Research's synthesized findings (Key Findings, Best Quotes, etc.) are no longer Serving Tray items — they're stored as full reports, listed under the Auto-Research run in the Sources rail, with the complete report action set (Open full, Share, Add to Tray, Delete, Copy, Download .md, Download PDF, citation toggle). They don't land in the tray automatically anymore; use the report's own Add to Tray button if you want a copy there — the Serving Tray has been a purely manual shelf since W1.14. If you're looking for an older run's extraction cards, find them under that run's row in the Sources rail rather than in the tray.

What is the Full Report option in Auto-Research?

Full Report (W1.18) is a box in the Auto-Research setup, right below the "What to EXTRACT" presets. When it's on, once the run finishes gathering sources and writing its extraction cards, the SAME background job keeps going and writes one complete written report on your topic, with citations, using the sources that run just gathered (no second gather, no separate Auto-Write run needed). It's built by the same engine as Auto-Write. When it finishes, the report shows up as a "Full report" card on the run in the Sources rail, opens in the reading pane, and gets the same actions as any report — Share, Add to Tray, Delete, Copy, Download .md, Download PDF — plus it's searchable in Ask Your Cafe like every other report type. As of W1.44, these actions sit in one row below the report body rather than a header strip, and gained Open full alongside the rest. As of W1.32B, it's on by default for every depth (it used to be off by default) — see Can I skip the report in Quick mode? above for how to turn it off on Solid/Deep.

Does Full Report cost more?

Yes. Full Report is an extra writing pass on top of the research itself, so having it on adds a second line to the cost estimate: "Research ~X + Report ~Y" — Y is the additional, depth-scaled cost of writing the report (the same hard-cap protections that apply to Auto-Write apply here too). On Solid/Deep, turning it off under Advanced → "Research only" costs nothing extra; Quick has no off switch, so a Quick run always includes Y. The affordability check covers X + Y together, so you won't be able to start a run you can't fully afford, report included.

What happens if my Full Report fails to write?

Your run still completes normally. Sources and extraction cards are saved to your Cafe as the run works, well before the report-writing step starts — so if the report itself fails partway through, nothing else is lost. You'll see a "Report failed" note on the run instead of a "Full report" card, but the run's status is still "complete," not an error. You can always re-run Auto-Write manually from your existing sources if you want another attempt at a full report.

Why is my Full Report about the wrong topic?

This used to trace to a gather-quality issue — occasionally the agent's Wikipedia search lane derived an unrelated search term from your topic (for example, a "fluoride toothpaste" topic pulling in nasal-polyp articles), and once that happened the off-topic sources could dominate the run. As of W1.18 Step 2, this is substantially fixed: the Wikipedia search term is now anchored to your actual topic, and every gathered source is tagged with a topical-relevance score — sources that score clearly off-topic are automatically excluded from what a report cites (they stay in your Cafe as raw, tagged sources; nothing is deleted). A borderline-relevant source, or one you explicitly pin into a report yourself, can still slip through — the relevance gate is the primary defense, not a hard guarantee — so it's still worth skimming the run's extraction cards (Key Findings / Summary) first as a sanity check before trusting a Full Report.

Do I need to know Markdown to write in the Workspace?

No. As of May 2026 (E70.1), the Workspace section editor is WYSIWYG (what you see is what you get). Type normally, and use the 15-button formatting toolbar at the top — Bold, Italic, H1/H2/H3, bullet / numbered lists, blockquote, inline / block code, link, table (with a hover-to-size grid picker), horizontal rule, undo, and redo. Keyboard shortcuts (Ctrl+B, Ctrl+I, Ctrl+Z, Ctrl+Y, Ctrl+S) work the way you'd expect from any modern editor. The old raw-Markdown textarea and "Preview" toggle are gone — there's nothing to preview, because what you see IS the preview.

What happened to the Preview button in the Workspace?

The Preview button is gone — replaced by the WYSIWYG editor itself. Because formatting now renders live as you type, there's no longer a separate "raw vs preview" mode. If you want to drop down to raw Markdown for any reason (pasting in pre-formatted content, copying syntax out, debugging an unusual layout), use the <> Source button in the section editor header to toggle into Source View. Click the pencil icon to switch back.

What is Source View in the Workspace?

Source View is a power-user mode that swaps the WYSIWYG editor for a plain Markdown <textarea>. Click the <> Source button in the section editor header to enter it; click the pencil icon to leave. Available to every signed-in user (no tier gate) — useful if you're pasting in pre-formatted Markdown, copying syntax out for use elsewhere, or just prefer typing raw. Your choice persists per-browser, so once you turn it on it stays on across sections and page reloads until you turn it off.

What's the "Ask AI about selection" button?

Drag-select any text in the WYSIWYG editor and a small floating menu appears next to your selection with an Ask AI about selection button. Click it and the Ask AI panel opens with the selected text pre-filled — a fast way to ask the AI to explain a passage, rewrite a sentence, or sanity-check a claim without leaving the editor and re-typing what you want to ask about.

How do I insert a table in the Workspace?

Click the table icon in the section editor toolbar. A small grid popover opens (10 columns × 8 rows) — hover over the cells to size your table (the picker shows you the dimensions live, e.g. "3 × 4"), then click to insert. Press Esc or click outside to cancel. Once inserted, click into any cell to type, and use Tab to move between cells.

Where can I see a list of every keyboard shortcut?

Click the ? button in the section editor header. A small popover opens listing every shortcut the editor supports: Bold, Italic, Undo, Redo, Save, discard, and Tab / Shift+Tab for list indent / outdent (or 2-space indent inside code blocks). The popover replaced the old "Markdown help" cheat sheet — there's no Markdown syntax to memorise any more, only the shortcuts you use day to day.

Does AI Write still type the draft out word-by-word?

Not in the new WYSIWYG editor. AI Write generates the complete draft on the server and inserts it into the editor in one step. You'll see a brief loading state, then the finished, formatted content appears all at once — ready to keep editing. The change makes the live-generation feel less like watching a stream and more like getting a finished draft back from a co-writer.

Where can I see the open-source software Coffeescribe uses?

The footer of every page now has an Open Source link (alongside Terms, Privacy, and AI Content). It opens /legal/open-source, which lists every open-source package Coffeescribe depends on along with its licence (MIT, BSD, Apache-2.0, etc.) and a link back to the project. This is our way of meeting the attribution obligations that open-source licences require — and a public thank-you to the projects we build on.

What are the AI Write modes?

The AI Write cockpit offers five writing modes:

• Continue writing — AI continues from where you left off
• Improve existing — AI rewrites and enhances your current text
• Write fresh — AI writes new content from scratch
• Write from research — AI writes based on your selected research notes
• Expand from research — AI expands your existing text using research notes

What is the Quality Score?

The Quality Score (0-100) is a system-calculated number that reflects how a scribe was created — not reader opinion. It's based on the AI model's writing ability, whether web research was used, the generation mode (slower = higher quality), verified content, freshness, and custom instructions. Quality scores appear on public scribes in the library, blurb page, and Scribe Hub. See Library & Sharing for the full breakdown.

What's the difference between Quality Score and User Rating?

Quality Score is calculated automatically by the system based on creation signals (model quality, research, generation mode, etc.). User Rating is the star rating submitted by readers. Both are shown on public scribes but measure different things — quality score reflects how the scribe was made, user rating reflects what readers think of it.

Does Coffeescribe check for plagiarism?

Yes. Every signed-in user can use the Check Plagiarism button in the Workspace to scan sections for potential plagiarism. It chunks your text into paragraphs and searches the web for matches. Note that AI-generated content is created on demand and is typically unique, but similar prompts may produce similar outputs. For high-stakes use cases (academic, publishing), consider using a dedicated third-party plagiarism service as well.

---

Writing Prompts

What is a Writing Prompt?

The Writing Prompt is the system-level instruction set the AI reads before it writes any scribe content — genre voice, structure, reader calibration, and style/tone, plus a locked structural contract CoffeeScribe always applies underneath (header fields, formatting rules) so nothing you edit can break generation. Coffeescribe now lets you see and edit it, either as the default for a category of future scribes (Settings → Writing Prompts) or as a one-off rewrite for a single scribe ("Edit writing prompt"). See Writing Prompts for the full walkthrough.

How do I edit the default writing instructions for my future scribes?

Go to Settings → Writing Prompts. Pick a Genre / Content type / Age range / Reader level / Style / Tone combination, and the four building blocks (Genre voice, Structure & shape, Reader calibration, Style & tone) resolve below it. Edit any block's text and click Save — it becomes the default CoffeeScribe uses for every future scribe matching that exact combination. Existing scribes are never affected.

Can I customize the writing prompt for just one scribe?

Yes. Once a scribe has a plan (its chapters and sections exist), open Edit writing prompt — a full button on the Structure Review screen, or a compact icon button in the Workspace section toolbar (beside the Web search toggle). It shows one big editable textarea with the scribe's complete effective prompt. Save your rewrite and it drives that scribe's introduction and every section — for new writes, regeneration, and resuming an interrupted scribe.

Does editing my Writing Prompts cost tokens?

No. Viewing, editing, saving, resetting, and reverting a Writing Prompt are all free — you're only editing instructions, not running the AI. Tokens are still spent exactly as before, when a scribe is actually generated using whichever prompt is in effect.

Is Writing Prompts a paid feature — do I need Pro or Creator?

No. Writing Prompts has no tier gate — Free, Pro, and Creator can all view and edit every block in Settings and every scribe's per-scribe prompt.

What's the difference between Custom Instructions and a Writing Prompt?

They're separate, additive layers. Custom Instructions (and per-section instructions in the Workspace) are short, discrete notes bolted onto a specific generation call — they always apply on top of whatever Writing Prompt is in effect. A Writing Prompt (Settings default or per-scribe edit) is the AI's underlying voice and structure instructions — the foundation Custom Instructions get layered onto, not a replacement for it. If a scribe has its own custom Writing Prompt saved, that one prompt wins over your Settings defaults for that scribe — but your Custom Instructions still apply alongside it either way.

Will editing a Writing Prompt change scribes I've already created or that are mid-generation?

No. A Settings block edit only affects scribes you create after saving it. A per-scribe prompt edit only affects that one scribe's future generation activity (new sections, regeneration, resume) — text already written stays exactly as it was written; nothing is retroactively rewritten.

Can I undo an edit to a Writing Prompt?

Yes, two ways. Reset to default (with a confirm prompt) discards a customization and returns to CoffeeScribe's stock wording. For a Settings block, History shows every version you've saved, newest first — click Use this version on any older entry to revert (this saves it as a new version; there's no separate "undo" step). Per-scribe prompts don't have version history — only Reset to default.

Does editing the "story" shape block turn citations back on?

No. Story-shaped scribes never show inline citation markers or a Sources block, regardless of what you write in the shape block — that rule is enforced by a separate step after generation, not by prompt wording. See Why doesn't my story show citations, even though I turned Web Search on? above.

---

Audioscribe

How do I create an audiobook?

Open your scribe and click the Audioscribe tab in the secondary navigation bar. Choose a voice provider and voice, confirm the credit cost, and click Generate. You can generate individual chapters or use Generate All / Generate Remaining for multiple chapters at once. Audioscribe is available to every signed-in user — token balance is the only limit.

What voice providers are available?

Coffeescribe currently offers two AI voice engines: OpenAI (high-quality, fast, 6 built-in voices) and Chatterbox (open-source, private, supports voice cloning). When both are available, tabs appear in the Generate Audio dialog so you can compare and choose. Each provider has a short description to help you decide.

Can I use my own voice?

Yes, using the Chatterbox provider. In the Generate Audio dialog, select the Chatterbox tab, expand "Record your voice," read the provided script, and upload your recording (WAV, MP3, or M4A). The AI will approximate your voice for narration. Voice cloning is only available with Chatterbox — OpenAI uses its built-in voices only.

How much does audiobook generation cost?

Cost depends on text length and the voice provider — Chatterbox is slightly cheaper than OpenAI for the same text. The exact cost is shown before you confirm so you know what you're spending. There's no monthly cap — generate as many audiobooks as your token balance covers, any time.

What happens if some chapters already have audio?

The button changes to Generate Remaining and only generates chapters that don't have completed audio yet. Your existing audio is never overwritten. To regenerate a specific chapter with a different voice, use the individual Generate button on that chapter.

Can I stop generation once it starts?

Yes. Click the red Stop button that appears while chapters are generating. Any chapters that have already completed keep their audio — only pending chapters are cancelled. You can regenerate cancelled chapters later by clicking Generate.

Does Generate All run faster now?

Yes. When you use Generate All, chapters are generated in parallel rather than one at a time. A 10-chapter scribe completes significantly faster than before.

Is Read Aloud different from Audioscribe?

Yes. Read Aloud is free, instant, and uses your browser's built-in speech. Audioscribe uses AI voice generation for higher quality, saves the audio permanently, and can be shared publicly. Read Aloud is available in Reading Mode; Audioscribe has its own dedicated tab in the secondary navigation bar.

Can I share an audiobook with someone who doesn't have a Coffeescribe account?

Yes. When your scribe is public, click Share on the audiobook page to copy a direct link. Anyone you send it to can open the link and play the audiobook — full playback controls (play/pause, seek, speed, chapter switching) all work without signing in. Actions that change anything (Download, Generate, Delete, privacy toggle) prompt the listener to sign in first. The link is stable and stays working even if you later hide individual chapters.

Will my audiobook resume where I left off?

Yes. If you pause or leave a chapter partway through, Audioscribe remembers where you stopped — per chapter, per your account. Come back on the same device or a different one and the player automatically seeks to your last position. You'll also see a "Resume from MM:SS" label and a thin progress bar on each chapter card showing how far through you got. Resume only works when you're signed in; anonymous listeners always start from the beginning. Your listening position is private and never visible to anyone else.

Can I generate an audiobook on my phone?

Playback on mobile is great. Generating on mobile is more fragile — if you lock your phone, switch tabs, or the screen times out during a long generation, the browser may throttle the background work and the UI can appear stuck. You'll see a soft warning banner reminding you that generation works best on a laptop or desktop. You can dismiss the banner and generate anyway, but for longer scribes we recommend starting the generation from a computer. If a mobile generation looks stalled, pull down to refresh — the visibility-aware refresh catches up any chapters that completed while the tab was backgrounded.

How long can my custom voice clip be?

Voice clips can be 10 seconds to 3 minutes (180 seconds) long, up to 25 MB in size. The best results come from 60-120 second clips — long enough to capture your tone and cadence, short enough to stay consistent. Clips longer than 180 seconds are rejected with a friendly message asking you to trim them. Clips under 10 seconds produce poor quality cloning. See the Audioscribe guide for full recording tips.

Does my audiobook link keep working if I make the scribe private?

Audiobook files are cached by the listener's browser and served from a public-read storage bucket, so anyone who already has the link may still reach the audio even after you flip individual chapters private. Treat sharing as permanent — only share with people you're happy to have ongoing access. If you need to revoke access to a specific generation, delete and regenerate it under a new URL.

I edited a section / combined sections / reordered things — does my audiobook update automatically?

No — the audiobook is generated per chapter, captured at the moment you ran Generate. It doesn't re-sync on its own when you edit. Three cases:

1. You reordered or combined sections within a chapter, or edited section text, or deleted a section — the chapter's audio still plays but now narrates the OLD text. Regenerate that chapter from the Audioscribe page to bring the narration back in sync with your edits.
2. You reordered whole chapters (dragged a chapter handle in the sidebar) — ✅ nothing to do. Each chapter's audio is unchanged; the player just plays them in the new order.
3. You only renamed a section title — the section title isn't read aloud (the audio narrates section content, not titles), so no regeneration needed.

Rule of thumb: changes inside a chapter leave that chapter's audio stale; changes between chapters are free. See Audioscribe → When edits in the Workspace affect existing audio for the full table.

---

Navigation

Where did the Notebooks tab go?

Notebooks has moved from the secondary navigation bar (inside individual scribes) to the primary navigation bar at the top of every page. This means you can access your highlights, notes, and verifications from anywhere in the app without needing to open a scribe first.

What is Audioscribe?

Audioscribe is the new name for the audiobook feature tab. You will find it in the secondary navigation bar when viewing a scribe. The feature itself is unchanged — it still lets you generate AI-narrated audio of your scribes, choose voice presets, or upload your own voice.

How do I know which scribe I am viewing?

The scribe title now appears directly in the secondary navigation bar, to the left of the tabs. This makes it easy to identify which scribe you are working with at a glance, without needing to scroll to the page header.

Where did the floating light/dark mode button go?

The theme button has moved from the bottom corner of the screen into the navigation bar at the top — look for the sun/moon icon next to your profile menu. Click it to cycle between light, dark, and system themes. It kept covering page content in its old floating spot. You can also still change the theme from the Theme section inside your profile menu. On pages without a navigation bar (like the landing page), your saved theme preference still applies — set it once from any page with a nav bar.

---

Library Access

Do I need an account to read scribes in the library?

Yes. The library, the Brew page, and individual scribe pages all require a Coffeescribe account. If you try to open any of these pages without being signed in, you will be redirected to the login page.

The only way to read a specific scribe without an account is via a share link sent to you by the scribe's owner. See below.

How do I share a scribe with someone who doesn't have a Coffeescribe account?

Open the scribe (or its blurb, audiobook, or Workspace page) and click the Share button. The link that gets copied to your clipboard includes a short code that lets the recipient read the scribe without signing in.

Do not just copy the URL from the address bar — that bare link will redirect a signed-out person to the login page. The Share button is the only way to generate a link that works for non-members.

Only public scribes can be shared this way. Private scribes redirect to login regardless of what link is used.

Why does a link someone shared with me ask me to sign in?

Two likely reasons:

1. The link is a bare URL, not a share link. The sender copied the address bar instead of using the Share button. Ask them to open the scribe, click Share, and send you the copied link — it will have a short code at the end (e.g. ?share=abc123).
2. The scribe was made private after the link was sent. Private scribes redirect all signed-out visitors to login, even with a valid share link. The owner would need to make the scribe public again.

If you have an account, you can also just sign in — signed-in members can read any public scribe directly without needing a share link.

---

Account & Sign-in

Do I need an invite code to sign up?

Yes, during private early access an invite code is required. Join the waitlist on the homepage to receive one, or ask an existing user for a spare code. See Getting Access above for full details.

How does the password reset flow work?

You request a reset on /forgot-password and Coffeescribe emails you a 6-digit code (not a clickable link). Type the code into the app, then set a new password. You're then redirected to /login to sign in fresh. The code expires 1 hour after it's sent. If you don't receive an email within a couple of minutes, check your spam folder, then click Resend code once the 30-second cooldown ends.

I didn't get my password reset email — what should I do?

Wait 1-2 minutes first — delivery is usually quick but can occasionally take a moment. Then:

• Check your spam, promotions, or "all mail" folder. Add noreply@coffeescribe.ai to your safe senders.
• Confirm you typed the email exactly as it is on your account (including any alias suffixes).
• If you originally signed up with Google or GitHub, you don't have a password to reset — sign in with that provider instead.
• After 30 seconds the Resend code button unlocks. Each new code invalidates the previous one — only the latest code in your inbox will work.
• If nothing arrives after 5 minutes and a couple of resends, use the in-app feedback button.

The reset code says it's invalid or expired — what's wrong?

A few common causes:

• Old code: Codes expire 1 hour after they're sent, and any earlier code is invalidated as soon as you click Resend. Make sure you're typing the code from the most recent email.
• Wrong digits: The field only accepts 6 digits — anything else is ignored. Re-type carefully.
• Rate limit: If you've requested several codes in a short window, Supabase may temporarily block new requests. Wait a couple of minutes and try again.
• Expired session: If you sat on the code-entry screen for over an hour, request a fresh code with Resend code.

Can I use a DuckDuckGo, Apple Hide My Email, or Fastmail alias?

Yes. Both signup confirmation and password reset use 6-digit codes that you type into the app, not clickable links — so URL-rewriting privacy forwarders can't break the flow. Just make sure you sign in with the same alias you signed up with, and that the alias is still active in your forwarder's dashboard so the email actually reaches you.

Do I need to verify my email address when I sign up?

Yes — if you sign up with email and password, Coffeescribe sends a 6-digit verification code to your inbox. Enter it on the verification screen to activate your account. Codes expire after 1 hour; request a new one if needed. Google and GitHub signups skip this step because the OAuth provider has already verified your address.

---

Bring Your Own API Key (BYOK)

Do I need to add my own API keys?

No. Coffeescribe provides all the provider keys you need to use the platform — BYOK is entirely optional. You only need to add your own key if you want your requests to a specific provider (OpenRouter, OpenAI, Apify, Mistral, OpenAlex, Google Books, or Venice) to draw from your own provider account instead of the platform's shared quota.

Do I need platform tokens if I use my own key?

Only as a backup. When you save a working OpenRouter key, generating scribes, Quick Reads, and Research does not check or deduct your platform token balance — the cost goes to your own provider account. Platform tokens are only spent if your key fails mid-generation and Coffeescribe falls back to the platform key to finish your work. If your key is working, generation is free of platform token cost.

What does adding my own API key actually do?

When you save your own key for a provider, two things happen: (1) your requests to that provider use your key, and the cost goes to your provider account; (2) the platform token gate is skipped — Coffeescribe does not check or deduct your platform balance for that generation. This means with a working OpenRouter key, you can generate scribes regardless of your platform token balance. Platform tokens are only used if your key fails and Coffeescribe needs to fall back.

Why did a generation use my platform tokens when I have my own key?

Your key failed mid-generation. When your key hits an error — an authentication failure, a revoked key, a per-key spend cap, a quota error, or any other provider-side rejection after retries — Coffeescribe automatically falls back to the platform key to finish your generation so your work isn't lost. A toast notification appears after the generation completes: "Your [Provider] key isn't working, so we used your tokens — fix it in Settings." Go to Settings → API Keys and click Test saved key to check if your key is still valid.

Can I test a key I already saved without re-pasting it?

Yes. In Settings → API Keys, once a key is configured the test button label changes to Test saved key. Click it without entering anything in the field — Coffeescribe tests your stored key server-side. A green toast means it is still valid; a red toast with an HTTP status code means it has been revoked, hit a limit, or has an issue. The plaintext key is never returned to your browser during this test.

What happens if my key runs out of credit?

If your key runs out of credit mid-generation, Coffeescribe detects the error and falls back automatically:

• If you have platform tokens: the generation finishes using platform tokens, and a toast appears linking you to Settings → API Keys to fix the key.
• If you have no platform tokens: the generation stops and you are shown a message with two options — top up your token balance, or fix or change your key in Settings → API Keys.

To avoid this, keep a credit balance in your provider account, or set a per-key spend cap above what you typically use per generation. A generation of a 10-chapter scribe on a mid-tier model typically costs a few cents on your OpenRouter account.

Does the automatic fallback work for audiobooks, Research Mode, and imports too?

Yes. The fallback isn't limited to scribe/Quick Read text generation — it covers every feature that uses a saved key: AudioScribe narration (OpenAI), Research Mode (Find the Truth, Auto-Write reports, claim verification, academic paper analysis, and Apify scraping/lookups), and Scribe Conversion OCR (Mistral) all fall back to the platform key the same way scribe generation does if your own key stops working mid-task.

Is my API key safe? Where is it stored?

Yes. Your key is encrypted with AES-256-GCM before it is stored in the database. After you save it, the plaintext key is never returned to your browser — you only ever see •••• last4 (the last four characters) as a confirmation. The key input is masked as you type. Only your account can access your saved keys (row-level security enforced in the database). We never log the plaintext key and do not share it with any third party.

Which features use which provider key?

You can mix and match — add only the keys for the features you use most. Any provider without a saved user key falls back to the platform key automatically.

Where do I manage my API keys?

Go to Settings → API Keys. All seven providers are listed. You can paste a key, test it live before saving (Test connection), save it (encrypted), remove it at any time, and test an already-saved key with Test saved key. A Manage API keys button also appears on the relevant feature surfaces (Create, AudioScribe, Research, Import) and opens a modal scoped to that surface's providers.

How do I get an API key for each provider?

Each provider card in Settings → API Keys has a "Where do I get this?" link that takes you directly to that provider's key-creation page. See Using your own API keys (BYOK) for a full table of all seven providers and their sign-up links.

---

Privacy & Data

Who owns the content I create?

You do. You retain ownership of your ideas, prompts, and edits. AI-generated content is licensed to you for any use, including commercial purposes.

Is my content private by default?

Yes — every scribe is private by default, for every signed-in user. Only you can see a new scribe until you choose to publish it to the public library.

Can I delete my scribes?

Yes. Every signed-in user can delete any of their scribes at any time, including completed ones.

How do I export my data?

Go to your Settings page and scroll to the Data & Privacy section. Click Export Data to download a ZIP file (coffeescribe-export-YYYY-MM-DD.zip) containing your profile, scribes (with metadata), highlights, and notes. This feature is available to every signed-in user and can be used once every 30 days.

How do I delete my account and data?

Before deleting, consider exporting your data first from the Data & Privacy section in Settings (see above). To delete your account, scroll to the Danger Zone section in Settings and click "Delete Account." You'll need to enter your password to confirm. The process permanently removes your account, personal data, and private scribes. Public scribes are re-attributed to "Deleted User" and remain in the library. Any remaining token balance is forfeited and cannot be transferred or refunded — token packs are one-time purchases, so there's no recurring subscription to cancel.

<!-- training-data tier distinction: owned by the privacy epic — do not edit in W39 -->

Does Coffeescribe train AI on my content?

• Free + Pro: Your prompts are sent to the CoffeeScribe Model (our curated default) via OpenRouter. Whether the underlying provider uses prompts for training depends on the specific model behind the brand label at the time. We pick the underlying model with this in mind.
• Creator: If you stay on the CoffeeScribe Model the same applies. If you pick a model from the picker, training behaviour depends on that provider's terms — see the picker's info icon for each model's details.

See our AI Content Policy for full details.

---

Citations & References

How do citations work?

When you generate a scribe with Web Search enabled, the AI cites its sources inline with small numbered markers like [1], [2]. Click a marker to see the source title, domain, and a link that opens the original page in a new tab. At the end of each chapter, a References section lists every source used in that chapter.

Can I edit or remove a source?

Not yet. Citations in v1 are read-only — you can't add, remove, or rewrite individual sources. Regenerating the section with Web Search on will produce a fresh set of citations. Manual citation editing may arrive in a future release.

Why doesn't my older scribe have clickable sources?

Citations only apply to sections generated or regenerated on or after 2026-04-14. Older scribes show [N] as plain text. To get live citations on an existing scribe, open the Workspace and regenerate the section with Web Search enabled.

Can I hide the inline markers while keeping the References list?

Yes. Open Reading Preferences (gear icon in the reader) and turn off Show inline [N] markers. Inline markers disappear, but the per-chapter References section still renders. The setting is saved to your browser.

Will citations appear in my PDF export?

Yes, by default. In the export modal, open Preview & Customise — the Include citations toggle is on by default. Inline markers render as superscript [N], and each chapter ends with a clickable References section.

Why doesn't my model cite inline with [N] markers?

Different AI models cite differently. Some (Perplexity Sonar/Sonar Pro, Gemini 2.0 Flash, Qwen 3 235B) reliably place numbered [1], [2] markers inline through the prose. Others (GPT-4o, GPT-4o-mini, Llama 4 Maverick, DeepSeek Chat, Grok 3 mini) return search sources but don't drop inline markers. CoffeeScribe adapts to whatever the model produced: when there are no inline markers, the chapter shows a "Sources consulted" bullet list at chapter end instead of a numbered list, with a one-line note explaining the gap. To get inline numbered citations, regenerate the section with a Perplexity model or Gemini 2.0 Flash. The model picker's info icon shows a Citation behaviour (web search on) line for each curated favourite so you know what to expect before generating.

What does the amber "Not in our recommended set" warning mean?

Every signed-in user gets the full model picker. It has three tiers of choice: CoffeeScribe Model recommended and pre-selected at the top, ~12 curated alternatives we've personally verified for writing quality and citation behaviour, then a collapsed advanced disclosure for the full 350+ OpenRouter catalogue. If you select a model from the advanced disclosure, the picker shows an amber "Not in our recommended set — quality and citation behaviour are unverified, and the model may fail to generate" note. The model is still available; the warning just means we haven't reviewed it. Some non-recommended models work fine; others (especially smaller models) may produce weaker prose, weaker citations, or fail to generate at all.

Why has my favourite model disappeared from Recommended?

We update the curated Recommended set when providers retire models or when a model's behaviour changes. For example, Anthropic retired the claude-3.5-sonnet endpoint on 2026-04-29, so we deactivated it and promoted claude-3.7-sonnet and claude-3.5-haiku into the curated set instead. Google's Gemini 2.5 Pro variants were demoted out of Recommended (they require reasoning mode, which doesn't fit our writing-focused picker contract) but remain available via Browse all for users who want them. Look for an amber "Not in our recommended set" note when picking from the wider catalogue.

A model can also drop out of the research recommendations on its own: if a provider stops offering it, it is removed automatically rather than left in the list, because a retired model is one you would pick and then get a failure from. It stays available in Browse all if the provider brings it back.

Why did I get "The selected model is no longer available"?

Occasionally a provider retires a model endpoint with little warning — when that happens, the model is still listed in your picker until the next sync, but generation requests against it fail. Rather than burning a generation cycle and showing an "Internal Server Error" toast, CoffeeScribe now checks the model's status before calling the provider and returns a clean "The selected model is no longer available — please pick a different model" message. No tokens are charged. Open the picker, pick a different model (one of the Recommended favourites is the safe choice), and re-run the generation.

What models are in the catalogue?

We curate the catalogue for writing scribes — the picker shows text-completion models suitable for prose. Vision-multimodal, image, audio, embedding, transcription, and code-specialised models are filtered out so they don't clutter your picker. If you specifically need one of those for your work, you'll need to use a different tool — they're outside the writing-scribe scope.

Exporting Your Scribes

Why does my PDF look like a book now?

Since W51, Quick Download PDFs (and HTML exports) are book-grade by default — a cover page, chapter opener pages with a small-caps "CHAPTER N" label and an accent rule, serif body text, and a clickable table of contents with PDF bookmarks. This applies to every signed-in user automatically, costs no tokens, and involves no AI — it's generated entirely by code from your scribe's existing chapters. See Exporting Your Scribes for the full walkthrough.

How do I get the old plain PDF?

Open Preview & Customise in the Export menu and switch the Theme (or preset) to Classic. Classic reproduces the original report-style layout exactly — no cover, no chapter openers, no accent colours. Preview & Customise is open to every signed-in user, no upgrade needed.

What is "Customise with AI"?

A third option in the Export menu (also available as an "Enhance with AI" button on the customise page) that asks an AI model to design your book's look: an accent colour, per-chapter eyebrow labels, optional italic chapter intros, optional verbatim pull-quotes, and cover copy. You can optionally type free-text design instructions (e.g. "warm academic palette, no pull-quotes") to steer it, or leave it blank for a fully automatic design.

Does the AI change my book's text or citations?

No. The AI design pass only chooses layout and cover/chapter decoration — it never rewrites your chapters or sections, never changes facts, and never drops citations. Any pull-quote it adds is always a verbatim excerpt of your own writing, checked automatically before it's used — the AI can't invent a quote.

Does Customise with AI cost tokens?

Yes — it's billed like any other AI feature, and it's owner-only (you can't run it on someone else's public scribe). Quick Download and the plain Preview & Customise page never use AI and stay free and private — only pressing "Customise with AI" / "Enhance with AI" sends your book's text to the model. Once run, the design is saved on your scribe, so re-downloading is free; pressing "Re-run AI design" bills again.

Why is there no page number in the footer?

The running footer at the bottom of each page shows your scribe's title rather than a page number — a current limitation of our PDF renderer when combined with the new book layout. Use the table of contents or your PDF viewer's bookmark/outline panel to navigate instead; a numbered-page footer is planned for a future release.

Feedback & Support

How do I report a bug or suggest an idea?

Click the Feedback button (bottom-left on every page). It opens with three choices — Report a bug, Suggest an improvement, or Something else. For bugs you'll get a short form (what happened, what you expected, optional steps, and how disruptive it is); for ideas, just tell us in your own words. We auto-detect which part of the app you're in so it reaches the right place. See Sending Feedback for the full walkthrough.

Can I send feedback anonymously?

Yes — tick Submit anonymously on the details step. Your feedback won't be tied to your name on any public lists, though our team can still act on it. (When you submit anonymously, the follow-up email option is hidden.)

Will you email me about my feedback?

Only if you let us. There's an "OK to email me follow-up questions about this" checkbox — leave it ticked and we may email you for more detail, or untick it to opt out. It's hidden for anonymous submissions.