Can I use Claude to extract the open questions from a brief, instead of just correcting answers?

Yes — that is what the question-extractor skill is for. Drop the SKILL.md into your Claude project, then tell Claude "turn the brief above into a questionnaire — JSON for passbackai.com." Claude reads the input (PRD, retro, email, transcript) and emits a JSON object with up to 12 open decisions, each with three or four concrete options. Paste that JSON into the textarea on this page and the site renders it as an interactive form — answer in 90 seconds, hit Copy, and you have a Markdown bundle to send back to whoever wrote the brief. The full launch note is at /blog/extract-open-questions-from-any-text.

Why does Claude keep rewriting the wrong paragraph when I ask it to fix something?

Because the chat box can carry your sentence but not your finger. When you type "fix the third bullet under Architecture — the first one, no, the second one," you are doing in prose what would take you half a second to do by pointing. The model has nothing to point at, so it picks a paragraph that sounds close and rewrites that. For one fix you can outwrite this. For five you can't — by correction four you stop describing which passage you mean, and the model fixes the wrong one because you stopped telling it. PassbackAI lets you highlight the actual passage and write the note next to it. The note ships with the verbatim quote AND the heading it sits under, so the model lands on one target instead of guessing between two.

How do I give ChatGPT feedback on multiple parts of its answer at once?

You stop typing it in the chat box. The chat box is fine for one correction — for five it costs you two of them. By comment three the input field has scrolled, the original answer has scrolled, and you are silently deciding that comment four "isn't worth it" and comment five "the model will probably catch on its own." It will not. The next answer comes back wrong on exactly those two points and you do another round. PassbackAI gives you somewhere to write all five comments while the answer stays put — highlight the passage, type the note, repeat — and copies the whole bundle out as one paste. All five land on the next turn. No round two for the two you almost gave up on.

Does PassbackAI work with Claude, ChatGPT, Cursor, or Gemini?

Yes — that's what it's for. Any chat that takes Markdown in and Markdown out is a fit. Paste the answer, highlight the five things that need work, copy it all back out as one Markdown block, paste that into the chat. You stop being the part of the loop that gets tired around comment four. The model sees five anchored corrections instead of three vague ones and a shrug.

How do I get Lovable or v0 to edit only what I named instead of regenerating the whole app?

Quote the exact live strings you want changed, write the edit beneath each, separate blocks with ---, and prefix the batch with "change only the literal strings I quoted; do not regenerate other components." That's the whole recipe. The model broadens by default because a vague note ("make checkout nicer") gives it no scope; a quoted string is an unambiguous target. Six edits land in one turn, and nothing you didn't mention moves. The per-tool walkthrough — with a worked example of eleven edits shipping in twelve minutes — is at /for/lovable.

How do I give Cursor feedback on a long Markdown file without it rewriting half the README?

Switch out of Composer or Agent and into chat mode, then use paired-passage feedback — quote the exact passage, write your note beneath it, separate blocks with three dashes. Cursor's editing tools are tuned for code, where a function is an unambiguous target; on a 1,200-word README they broaden context and hand back a diff of 240 lines. Chat mode plus the paired format turns the same model into a surgical editor. The full Cursor-specific playbook is at /blog/give-cursor-feedback-long-markdown-file.

Does the model actually read the structured comments?

The export format is designed to be unambiguous to a language model. Each comment appears as a quoted passage in curly quotation marks followed by your note, with --- separators between blocks. The model parses this as paired data without prose interpretation. On the next turn, you can see exactly which comments landed and which the model skipped — it turns a coin-flip review loop into a deterministic one. Each quote also carries the heading it sits under — or, for tables, its row or column — so the model lands every fix on its exact target.

Do you store my document anywhere?

No. There is no backend. No database. No "we may use your content to improve our service." Your document lives in your browser's localStorage until you clear it — it never leaves the browser. (The site loads anonymous page-view analytics from Vercel; your document is never part of those pings.)

Do I need an account?

No. No email, no password, no "sign in with Google." The privacy policy is that there is nothing to have a policy about.

Is sharing private, or does it need a signup?

Sharing needs no signup, and it stays private. When you share a document — a feedback session or a questionnaire, empty or already worked-on — the whole thing is encoded inside the link itself, in the URL hash fragment. The browser never sends that fragment to a server, to analytics, or in the Referer header, and there is deliberately no backend to upload it to. So a shared document is never stored on a server: it lives only inside the link you send, and only the people who hold that link can open it. No account on either end — the sender doesn't have one, and the colleague who opens the link doesn't need one.

Can I work on more than one document at once?

Yes. PassbackAI keeps several documents open at once — a marked-up Claude answer, a questionnaire from a PRD, a second draft you're still reviewing — and you switch between them from the documents sidebar without losing any. Each one waits locally in your browser while you work on it (nothing is uploaded), and each can be shared as its own self-contained link.

Fix an AI answer in one pass — paste the draft, mark the fixes, pass it back.

A long reply lands from Claude, ChatGPT, Cursor, or Gemini. Most of it is right — a few lines aren't. PassbackAI opens it as a real document you can read, mark up, and send straight back the moment you spot what's off.

Chat box vs PassbackAI

The same five fixes, two ways. (The right column carries one comment — scoped to the whole column, not retyped per cell.)

The job	The chat box	PassbackAI
Send five fixes at once	you give up after three	five notes, one paste
Point at the exact line	one text field — you can't	every note anchored to its words
Loop in a teammate	paste elsewhere, lose the thread	one link, no sign-up
Keep it private	it lives on a server	nothing leaves your browser

1 · Mark every fix where it lives

Highlight a line, drop a note, keep reading. Each note stays pinned to the exact words — and remembers the heading above it, so the model learns what to change and where. The tenth fix is as cheap as the first.

Take a sentence the model oversold:

Per my previous correspondence, kindly be advised the deliverable shall be furnished by end of week.

One Delete label, struck through — no retyping "cut this" down a long page.

2 · Send it all back in one paste

Hit Copy and every note bundles with its quote into a single block your model can act on — nothing scattered across a chat, nothing lost:

## Delivery
> "...furnished by end of week."
  [Delete] Too stiff — say "I'll send it Friday."

> the deliverable shall be furnished
  Give a real date, not "end of week".

3 · Or hand it to a teammate — one link, no sign-up

Need a human in the loop? Share the whole document as a single link — empty for a colleague to fill in, or already marked up. No account for you, none for them.

One click turns the doc into one link. No sign-up, ever — for anyone.
Send it blank or fully commented; they open it and keep going right from your marks.
Sensitive? Add a password and the link unlocks only for the people you tell.

4 · 100% private — by design, not by policy

There is no upload, no database, no log. The entire document rides inside the link itself, and a password encrypts it end to end. Close the tab and nothing is left behind — because nothing was ever sent anywhere.

_The marks above are live. Highlight any sentence on this page and a note opens right where you marked — that's the whole tool, working on itself._

Paste your own answer Make AI ask first (Beta)