https://passbackai.com/blog A field manual for giving LLMs precise, multi-point feedback. Diagnoses, playbooks, and per-tool recipes for Claude, ChatGPT, Cursor, Gemini, and AI-builder surfaces like Lovable. en-us © 2026 Elad Diamant Sun, 12 Jul 2026 00:00:00 GMT https://passbackai.com/og.png https://passbackai.com/blog https://passbackai.com/blog/interactive-decision-widgets-inside-a-document https://passbackai.com/blog/interactive-decision-widgets-inside-a-document Sun, 12 Jul 2026 00:00:00 GMT Inside a PassbackAI document your AI can embed decision widgets, one per unclear point: single choice, multi choice, open question, prioritize, and questionnaire. The real subject is the humane design that makes them feel like an assistant that leans without deciding for you — a Recommended badge that marks an option but never pre-selects it, an automatic Other row and a one-tap pencil so you're never caged in the AI's list, a Note field where an explanation alone is a valid answer, a one-line context that says why each question is open, and a prioritize block that reports back even when you change nothing, because 'I looked and I agree' is a real signal. Widgets are woven inline where each decision arises; settled thinking stays prose you can still comment on. RTL/Hebrew is automatic and the palette is growing. https://passbackai.com/blog/second-opinion-on-an-ai-answer https://passbackai.com/blog/second-opinion-on-an-ai-answer Thu, 09 Jul 2026 00:00:00 GMT The AI answered with confidence you can't verify — a contract clause, a severance analysis, a pricing plan — but you know the person who could check it in ninety seconds. Today that hand-off happens as four screenshots into WhatsApp, a formatting-destroying paste, or a shared-chat transcript your friend can read but not mark — and the feedback comes back detached from the text it's about. The fix is two clicks: paste the answer into PassbackAI, hit Share, hit Copy link, and send it wherever you already talk. Your friend opens a clean document in any browser — no account, no app, nothing uploaded (the link carries the document) — and pins signed notes to the exact sentences. Then the whole marked-up answer goes back to your AI in one paste. https://passbackai.com/blog/give-claude-feedback-without-copy-paste https://passbackai.com/blog/give-claude-feedback-without-copy-paste Thu, 09 Jul 2026 00:00:00 GMT Six rounds into a plan with Claude, you notice your real role: courier. Copy the draft out, paste it into the review tool, mark it up, copy the feedback, paste it back. Connect PassbackAI as a Claude connector (one URL, one OAuth) and both legs disappear: Claude routes its draft with route_document, it opens at a review link as a real rendered document, you pin notes to the exact sentences — and answer the questions Claude wove into the prose, one widget per genuinely unclear point (single choice, multi choice, open question, prioritize, questionnaire). Send back submits the review; Claude reads it with list_responses as anchored annotations plus structured answers — typed decisions, not vibes. The /passback skill teaches Claude to ask first. Honest edge: routing is the deliberate, opt-in exception to local-only — server-stored so Claude can read the responses, revocable for good. https://passbackai.com/blog/work-on-multiple-ai-drafts-at-once https://passbackai.com/blog/work-on-multiple-ai-drafts-at-once Wed, 08 Jul 2026 00:00:00 GMT A report, an email, and a spec, each buried in its own Claude or ChatGPT thread. PassbackAI keeps your recent AI drafts on a bench — one shelf you switch across from the top bar — so being unfinished with one never blocks starting the next. There's no finishing-tax: nothing has to be closed or completed first. The bench holds the last ~20 drafts (the least-recently-touched rolls off, never the one you're on), buckets them into Pinned / Today / Yesterday / This week / Earlier, and lets you pin the one you can't lose so it can't roll off. Paste a new version of a draft already on the bench and the Auto-Diff auto-detects the match and paints exactly what the AI changed between the two versions — the only convenient, visual way to see what the model actually touched. It all stays in your browser. https://passbackai.com/blog/feedback-for-ai-builders https://passbackai.com/blog/feedback-for-ai-builders Tue, 07 Jul 2026 00:00:00 GMT AI builders rebuild from your prompt, not from a diff — so every edit re-rolls the neighborhood, and the one line you wanted changed comes back with three components you never touched. The escape hatch is one move that ports across every builder: paste the rendered output or code, mark the exact spans that are wrong with a note on each, and pass the anchored batch back in one shot so it edits those lines instead of regenerating from your description. The pattern, plus a short per-tool note for Vercel v0, Replit, Base44, Bolt.new, Sticklight, and Lovable. https://passbackai.com/blog/what-is-passbackai https://passbackai.com/blog/what-is-passbackai Tue, 07 Jul 2026 00:00:00 GMT A 3-minute orientation primer. PassbackAI is the review-and-passback loop for AI drafts: when a model hands back a long answer with several specific things wrong in specific places, you paste it, mark each one where it sits, and pass the whole batch back in one anchored paste — every note landing on its exact target. It's what turns solo AI work from vague nudging into precise, professional editing, and it works across Claude, ChatGPT, Cursor, and Gemini (Claude the deepest, over MCP). Covers the core loop, cross-platform reach, when to use it and when not to, where your document goes, and the two ways in — you paste, or your AI routes it to you. https://passbackai.com/blog/delta-principle https://passbackai.com/blog/delta-principle Mon, 06 Jul 2026 00:00:00 GMT In a growing class of AI-native work, the draft is not the durable unit — the reusable unit is the delta it helps reveal: a scoped, anchored update to the model's working picture of your intent. Drafts are ephemeral; deltas accumulate. The essay lays out the Delta Principle, a falsifiable test (the Regeneration Criterion: a delta log earns its keep when a fraction of the tokens buys almost all of the outcome), six boundary cases where the principle does not hold, and why today's platform memory features are exhaust with a UI. https://passbackai.com/blog/share-a-doc-without-an-account https://passbackai.com/blog/share-a-doc-without-an-account Tue, 09 Jun 2026 00:00:00 GMT PassbackAI has no accounts, no logins, and no server — and it has a Share button anyway. The trick: the link doesn't point at your document, it carries it, compressed into the URL's hash fragment — the part of a URL the browser never sends anywhere. Nothing uploaded, nothing stored, nobody signs in on either end. Add a password and the same payload travels end-to-end encrypted, with the password never in the link. The recipient adds their own anchored notes and hands the draft back — and the merged feedback returns to the model in one paste. https://passbackai.com/blog/review-ai-output-takes-too-long https://passbackai.com/blog/review-ai-output-takes-too-long Thu, 04 Jun 2026 00:00:00 GMT Reviewing a long AI answer is real work, so most of us do less of it than we mean to — one tired read, two or three fixes marked, the rest waved through. But every issue you skip comes back next turn as a still-wrong answer, which buys another full generation: more tokens, more latency, another wall to re-read. The expensive part isn't the model rerunning; it's that the chat box makes a complete read so costly that you under-review. Make the read cheap — a real rendered document, a contents rail, an auto-diff — and one thorough pass replaces three half-passes. https://passbackai.com/blog/extract-open-questions-from-any-text https://passbackai.com/blog/extract-open-questions-from-any-text Thu, 14 May 2026 00:00:00 GMT 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: a choice, a ranking, an open answer. Paste it into passbackai.com and answer in 90 seconds, or forward it to whoever needs to decide. https://passbackai.com/blog/omgfixmd-chrome-extension https://passbackai.com/blog/omgfixmd-chrome-extension Thu, 30 Apr 2026 00:00:00 GMT The PassbackAI Chrome extension intercepts Copy on Claude, ChatGPT, Gemini, and Perplexity and shows a small "Copied. Review in PassbackAI →" toast. One click opens passbackai.com with the answer pre-loaded — the six steps between Copy and reviewing become one click. Runs in your browser, no telemetry — the copied content never leaves your browser. https://passbackai.com/blog/give-cursor-feedback-long-markdown-file https://passbackai.com/blog/give-cursor-feedback-long-markdown-file Fri, 24 Apr 2026 00:00:00 GMT Cursor's tools are tuned for code, where ambiguity is rare. Long Markdown prose has more ambiguity per line — and Cursor edits it like code, which is why the README keeps coming back rewritten. The playbook for keeping prose edits scoped. https://passbackai.com/blog/prompt-to-edit-only-specific-passages https://passbackai.com/blog/prompt-to-edit-only-specific-passages Fri, 24 Apr 2026 00:00:00 GMT Every "only edit X, leave the rest alone" prompt ranked by reliability, with the honest answer about why none of them work on their own — and the paired-passage format that does. Copy-pasteable templates included. https://passbackai.com/blog/chatgpt-rewrites-whole-document-when-i-ask-for-edits https://passbackai.com/blog/chatgpt-rewrites-whole-document-when-i-ask-for-edits Fri, 24 Apr 2026 00:00:00 GMT You asked for one small edit. It came back as a different document. Not your prompt — the chat box has no target selection, so every ask reopens the whole file. Here's what's actually happening and the format that stops it. https://passbackai.com/manifesto https://passbackai.com/manifesto Tue, 21 Apr 2026 00:00:00 GMT An AI draft is worth everything for four minutes, then it's spent. PassbackAI is the station those minutes happen in: paste the answer, mark every passage that needs work, and pass the whole batch back to your AI in one round-trip. A manifesto on why drafts belong in transit, not an archive. https://passbackai.com/blog/how-to-comment-on-markdown https://passbackai.com/blog/how-to-comment-on-markdown Tue, 21 Apr 2026 00:00:00 GMT Markdown has no native comment syntax. Every practical method — HTML comments, Google Docs, Notion, GitHub PR, Slack, purpose-built comment layers — ranked, with the 2026 case that matters most: giving the file back to the language model that wrote it. https://passbackai.com/blog/give-llm-feedback-multiple-passages https://passbackai.com/blog/give-llm-feedback-multiple-passages Tue, 21 Apr 2026 00:00:00 GMT The chat box loses at multi-point feedback. Every LLM user hits this wall around comment five. Here is the pattern that actually works — paired passages with notes — with a manual recipe for anyone, and a tool for anyone tired of doing it by hand.