Guides for giving your AI precise feedback
Diagnoses, playbooks, and per-tool recipes for the moment typing corrections into the chat stops working — and the review-and-passback loop that lands all of them in one turn, across Claude, ChatGPT, Cursor, and Gemini. Each piece stands alone, so start anywhere. If you want the short version of why the chat box loses at this, the manifesto is it. Or subscribe via RSS →
What is PassbackAI?
New here? Start with the 3-minute orientation. What the tool is, the paste → mark → pass-back loop, the AIs it works with, and where your document goes — the whole picture before you dive into the how-tos.
The GuideHow to give Claude or ChatGPT feedback on 5 things at once
The paired-passage pattern — a manual recipe you can run in any plain-text editor, and the mechanism for why structured input beats prose every time.
Why it breaksWhy ChatGPT rewrites the whole document when you only asked for one fix
The chat box has no target selection. RLHF trained the model to return complete documents. Four prompt attempts that almost work, one format that does.
Every incomplete AI review costs you another generation
Reviewing a long answer is so expensive in a chat box that you skim once and skip half the fixes — and every fix you skip comes back as another full generation. Make the read cheap and one pass replaces three.
How to get a second opinion on an AI answer (two clicks, no screenshots)
The AI sounds sure; your friend would actually know. Skip the four-screenshot hand-off: share the answer itself as one link — no account on either end — and get feedback pinned to the exact sentences, ready to pass back to the AI.
A prompt that gets Claude to only edit the passages you named
Every "only edit X" variant ranked by reliability, with copy-pasteable templates for Claude, ChatGPT, Cursor, and Gemini — plus a system-prompt rule for API and Project users.
How to add comments to a Markdown file (2026 edition)
Every practical method — HTML comments, Google Docs, Notion, GitHub PR, Slack, purpose-built comment layers — ranked, with the 2026 case the old methods don't cover.
Extract the open questions from any brief, as an interactive document
A Claude skill that reads any input — PRD, brief, retro, email, transcript — and weaves up to 12 open decisions into one interactive document, each point wearing the widget that fits it. Answer in 90 seconds, or forward it to whoever needs to decide.
How to give Claude feedback without copy-paste — the MCP loop
Connect once and the carrying disappears: Claude routes the draft to a review link, you mark the exact sentences and click through the questions it wove into the prose, and Claude reads the answers back itself — as data, not vibes.
How to give AI builders feedback on the exact lines
v0, Replit, Base44, Bolt.new, Sticklight, Lovable — they all rebuild from the prompt, not from a diff, so one edit re-rolls the neighborhood. The one move that stops it: point at the exact lines instead of re-describing them.
How to give Cursor feedback on a long Markdown file
Cursor edits prose like code: broadens context, applies a coherent diff, regenerates the neighborhood. Composer vs Agent vs chat, with a .cursorrules recipe that makes scoped edits the default.
How your AI asks you to decide — without taking the decision away from you
The interactive widgets your AI can drop into a document, right where each fork is — choose one, choose many, write, order, group. The craft is the restraint: a Recommended badge that leans but never picks for you, an Other row so you're never caged in the list, and a note that counts as an answer.
How to work on several AI drafts at once
A report, an email, and a spec, each buried in its own chat thread. The bench keeps your recent drafts on one shelf — switch from the top bar, pin the one you can't lose, and diff exactly what the AI changed between two versions. Unfinished isn't blocked.
The Chrome extension that gets your copied AI answer into PassbackAI in one click
The PassbackAI Chrome extension intercepts Copy on any of the four and shows a toast. One click forwards the answer to passbackai.com pre-loaded. Six steps become one. Runs in your browser, no telemetry.
How do you share a doc when nobody has an account?
No server, no logins — and a Share button anyway. The link doesn't point at your document; it is your document, riding in the part of a URL the browser never sends anywhere. Optional password makes it end-to-end encrypted.
We're Saving the Wrong Thing: The Delta Principle
Drafts are ephemeral; deltas accumulate. In a growing class of AI-native work, the draft is not the durable unit — the reusable unit is what it taught the model about your intent. The Delta Principle, a falsifiable test, and six places it breaks.
So much of what we read now is AI text with a short shelf-life — useful for one session, then gone. I wanted a faster, sharper way to work on it — and not alone.
So: mark what's wrong, keep your last 20 a click away, and when you want another set of eyes, hand the whole thing off as one link — your comments ride along inside it. Nothing is ever uploaded; it all stays on your device.
Made by Elad · LinkedIn
Read the full story →