# Someone sent you a document to review. _You can hand the whole job to your AI._ A PitchStation review link — a contract, a proposal, a webpage — lets you comment directly on the document: select text to object to it, propose replacement wording, pin visual issues. No account, no signup. And if you work with an AI assistant — Claude Code or Codex in a terminal, or plain claude.ai / ChatGPT in the browser — you can hand it the link and let it do the reading and the commenting **as you**. This page shows every path. ## §01 The 30-second version **Copy the full share link** The whole URL you were sent, including everything after the `?` — that part is your access key. **Paste this prompt into your AI** Fill in the two highlighted parts. Works in Claude Code, Codex CLI, Cursor, or any agent that can make HTTP requests. **Using claude.ai or ChatGPT in the browser?** Use the prompt in §03 instead — same result, one extra paste. **Open the link and check the Threads panel** Everything your AI posted is there under your name. Reply, refine, or delete anything — it's your review. **The prompt** ``` Review this document and leave comments on my behalf: Fetch the page first — near the end of its HTML is a note for AI agents pointing at a review-guide URL. Fetch that guide and follow it: it explains how to read the document and post anchored comments through the review API. Sign every comment as: When you're done, list what you posted and anything you chose not to post. ``` That's genuinely all a recipient needs. The rest of this page explains what happens, what your AI can and cannot do, and how to review by hand. [Diagram: Four ways to review — by hand, a terminal agent, a browser chatbot, or an extension — all landing as comments on one shared document.] ## §02 What your AI actually does Every review-enabled share is **self-describing to AI agents**. The page carries a machine-readable pointer to a guide at `/s//review-guide` — protected by the same access key as the document itself — which teaches the agent the full loop: - **Read the document.** The share URL returns the real HTML (for a multi-file bundle, each file is listed and readable individually). - **Identify as you.** Comments need a human name behind them. The agent registers the name you gave it once; every comment it posts is attributed to that name, marked `(unverified)` so the owner knows it wasn't a signed-in account. - **Post comments anchored to exact text.** A comment on a clause quotes that clause verbatim. The server **verifies the quote against the document** before accepting it — if the AI paraphrases, the comment is rejected with instructions to re-read and copy exactly, so a sloppy quote can never produce a comment that points at nothing. - **Propose replacement wording.** For contracts, the strongest move: a _suggested edit_ shows the current wording and the proposed wording side by side, and the owner can accept or reject it with one click. - **Read the discussion.** The agent can poll the thread to see the owner's replies and the status of each point (open / resolved / accepted / rejected). 🔒 **Your AI can't do anything you couldn't do by hand.** It uses the same access key you were sent, the same comment API the on-page widget uses, and the same rate limits. If the link is revoked or expires, both of you lose access together. ## §03 Using web claude.ai or ChatGPT (no terminal) Browser chatbots can **read** the shared page (their browsing tools fetch URLs) but cannot **post** — they have no way to call the comment API. So the flow has one extra step: the AI writes the review as a structured block, and **you paste it into the page**, where every quote is verified and nothing publishes until you confirm. **Give the chatbot this prompt** It reads the document with its browsing tool and produces a `pitchstation-review` block. **Copy the block it produces** One fenced code block containing all its comments. **Paste it into the page** Open the share link → **?** in the review panel → **“Paste an AI review”** → paste → check the preview → **Post**. Comments the page can’t find are flagged and skipped, never guessed. **Web chatbot prompt** ```` Read this document with your browsing tool: Also read /s//review-guide?k= — the section "If you cannot make POST requests" defines the output format. Review the document as my representative. My position: You cannot post directly. Output your entire review as ONE ```pitchstation-review``` code block per the guide — quote every passage VERBATIM (the page verifies quotes and skips anything it can't find), one point per comment. I will paste the block into the page myself. ```` ✅ **This flow has a safety property the direct one doesn't:** you see every comment in a preview, on the page, before anything posts. For a contract review where the comments speak for your side, that checkpoint is worth the extra paste. 🤖 **Agentic browser modes skip the paste.** ChatGPT's agent mode and Claude's browser extension drive a real browser — they can open the link and use the review UI directly: switch to Comment mode, select text, click to pin. For those, the §01 prompt works as-is. ## §04 Reviewing a contract with AI For a contract or agreement, tell the AI what your position is — it reviews on your behalf, so it needs your instructions, not just the document: **Contract review prompt** ``` Review this contract on my behalf: Fetch the page, find the AI review-guide note in its HTML, and follow it. Sign comments as: My position: For each clause that conflicts with my position, post a SUGGESTED EDIT (quote the exact current wording, propose the replacement wording, and give a one-sentence reason). For anything merely worth discussing, post a plain comment. One point per comment. Don't post anything about clauses that are fine — summarise those for me instead. ``` ### What happens across revision rounds When the owner revises the document, it stays at **the same link** and your comments come along: | If the owner… | Your comment… | | --- | --- | | kept the text you quoted | stays anchored to it, exactly where it was | | revised the wording you flagged | is marked **✓ addressed** — they acted on your point; confirm rather than re-argue it | | deleted the text entirely | is marked **⚠ orphaned**, with your original quote preserved — the point was never silently re-attached to different words | | accepted / rejected your suggestion | shows that decision, with their reason, in the thread | Your AI can read the changes too: `/s//changes` (same access key) returns every round’s note and a redline — `- removed` / `+ added` paragraphs — between rounds, so a returning agent knows exactly what moved without re-reading the whole document. The machine guide points agents at it automatically. Returning to the link after a revision shows an _"updated since you read it"_ banner with the owner's note about what changed — a natural moment to run your AI over it again: _"the document was revised — re-read it, check which of my points were addressed, and comment on anything new."_ [Diagram: Returning after a revision: compare versions, read the /changes redline, then act only on what actually moved.] ## §05 Reviewing a webpage or design Shared webpages (including multi-page sites shared as a bundle) take the same comments, plus visual ones: - **Numbered pins** — in Annotate mode, click any element (a button, an image, a heading) to drop a numbered marker and comment on that spot. - **Drawn boxes** — drag to outline a region ("too much whitespace here"). - **Viewport context** — every comment records the screen width it was made at, so "the menu is broken" made on a phone is understood as mobile-layout feedback. - **Per-page threads** — on a multi-file bundle, comments attach to the page they're about. **Design review prompt** ``` Review this website before launch, on my behalf: Fetch the page, find the AI review-guide note in its HTML, and follow it. This is a multi-page bundle — review every page, and put each comment on the page it belongs to (the guide explains fileId). Sign comments as: Look for: broken or placeholder content, claims that overpromise, unclear calls to action, inconsistent naming or pricing, missing legal basics (privacy, contact), and copy that reads machine-written. Post one comment per issue, quoting the exact text. Summarise anything structural that a comment can't capture. ``` 💡 **Where the AI can't click, you can.** An agent reads the page's HTML — it won't experience hover states, animations, or how the layout feels on your phone. The strongest review is both: let the AI sweep the content, then spend five minutes in Annotate mode pinning what only a human eye catches. ## §06 Reviewing by hand Open the link and look for the mode bar in the bottom-right corner. The **?** button in the panel header shows this same cheat-sheet — plus a **"Copy the AI prompt"** button that produces §01's prompt with the link already filled in, and **"Paste an AI review"** for the §03 flow. **View** Read normally. Nothing you do here is recorded. **Comment** Select any text to comment on it or propose replacement wording. **Annotate** Click anything to drop a numbered pin; drag to draw a box around a region. **Threads** Every comment — yours and others' — with replies and statuses. The first time you comment you'll be asked for a name. That's the entire signup. On a translated view (the Translate pill), the widget will ask you to switch back to the original before commenting — comments anchor to the original text so they survive revisions. ## §07 Good to know - **The owner sees everything in one feed** — your comments, your AI's comments, and other reviewers', each attributed. There is no private draft state: posted means visible. - **Check what your AI posted.** It signs your name. Open Threads after it runs; you can edit or delete your own comments (same browser/session — your identity lives in a cookie the AI's session also holds if it kept its cookie jar). - **Rate limits are enforced** per reviewer per document. If your AI hits one, it should stop and tell you rather than retry in a loop — the guide instructs it to. - **Don't paste the link anywhere public.** The URL contains the access key; anyone holding it can read the document and comment. Treat it like the document itself. - You can review in your own language — comments may be written in any language; only quoted text must match the original document. - **Plain text only** in comments (up to 4,000 characters). Markdown isn't rendered. ## §08 For senders: enabling this Turn on document review when you publish (or later — same link): | Surface | How | | --- | --- | | Share composer | tick **“Let viewers comment”** when creating the share | | MCP / AI assistant | `pitchstation_publish` with `comments: true` — or add `comments: true` to a `pitchstation_replace` round | | API | `POST /api/share` with `"comments": true` — for any client that can make an HTTP request | ### The owner’s AI loop: gather, draft, decide You don’t read the feedback by hand either. On your share’s analytics page, the Review section has a **🤖 Use your AI** button with a ready prompt for whichever AI you use — each one reads the comments, drafts the revision, and keeps you in control of what publishes: | Your AI | How the loop runs | | --- | --- | | Claude Code / Codex / Kimi (terminal) | Reads the feed (`pitchstation_comments`), drafts the revision and a per-thread decision list, **shows you both**, then on your go-ahead publishes (`pitchstation_replace`, announce) and records every decision (`pitchstation_comment_resolve`). Same link, comments re-anchor. | | claude.ai / ChatGPT / Gemini (browser) | Reads the document and the comment feed with its browsing tool, outputs the revised HTML plus a decision list. You apply: **Replace** on your Links page uploads the new version; the accept/reject/resolve buttons on the Review tab record the decisions. | | Browser-extension agent | Operates your logged-in session: reads the Review tab, narrates each decision before clicking it, and applies the revision through the Replace modal — you supervise in real time. | [Diagram: A full negotiation, assistant to assistant: each side’s AI works its half, meeting in one audited thread.] 🔁 **A full negotiation can run assistant-to-assistant.** Their AI reviews and objects; your AI reads the objections and drafts round 2; every exchange lands in the same audited thread with a human name behind it — and both principals see everything before it publishes. Then tell your recipient one sentence: _“Comment directly on the page — or if you use an AI assistant, paste it the link and ask it to review; it will find the instructions itself. Guide: pitchstation.ai/ai-review-guide.html”_. Your own AI reads the same discussion via `pitchstation_comments` and publishes revisions with `pitchstation_replace` — so a full negotiation can run assistant-to-assistant, on the record, with both principals in control.