How to get feedback on an AI-generated prototype
An assistant can produce a clickable prototype in minutes. Getting useful feedback on it takes longer. A screenshot loses the interactions, a zip file never gets opened, and "looks good 👍" in chat doesn't tell anyone what to change. Here's a workflow that gets specific feedback and feeds it back to the agent.
1. Make it one self-contained HTML file
Reviewers should open a link and see the real thing. That works when the prototype is a single HTML file with its CSS, JavaScript and images inlined, or loaded from absolute https:// URLs.
- If the assistant wrote one file, you're done.
- If it built a React or Vue app, have the bundler emit one file. With Vite that's
vite-plugin-singlefile. See single-file HTML builds. - Avoid relative paths like
src="assets/logo.png". Nothing else gets uploaded with the file, so they won't load.
2. Share a live link, not a picture
Upload the file to Ollav: drag it onto https://app.ollav.app, or let your agent do it with the share_html MCP tool or one curl call:
curl -F "file=@prototype.html" -H "Authorization: Bearer $OLLAV_API_KEY" -H "X-Ollav-Client: agent" https://app.ollav.app/api/documentsYou get a stable link. Workspace members can open it straight away, and you can share a single document with an outside reviewer as a guest.
3. Ask for feedback on elements, not in a thread
Ask reviewers to click the thing they're talking about and comment there. Each comment is pinned to that element, so "this" always means something. Make the request specific:
- "Comment on anything that would stop you finishing checkout."
- "React 👍 on comments you agree with instead of repeating them."
- "Resolve your own comment if a later version fixes it."
Reactions keep the noise down: five people agreeing is one comment with 👍 5, not five replies.
4. Feed the comments back to the agent
The agent that built the prototype can read the feedback directly. get_review returns the open threads as Markdown, each with the text of the element it's pinned to, so the agent can find it in the source:
Read the open comments on the Ollav prototype, fix them, publish a new version
and resolve what you fixed.Under the hood that's get_review → edit → publish_new_version → resolve_comment. The link stays the same and earlier comments are kept and marked as being on an older version.
5. Repeat until it's quiet
Each round, reviewers reload the same link. When get_review comes back empty, you're done. The link and the full discussion stay with the work, for whoever picks it up next.
Checklist
- One HTML file, everything inlined or absolute
- Shared as a live link with the right visibility (workspace by default)
- Reviewers comment on elements and react instead of repeating
- The agent reads open threads, ships a new version, resolves what it fixed