In this guide
Why screenshots and chat messages fall short
When an AI coding agent builds your site, the agent only knows the code. You only see the rendered page. A sentence like “the button on the pricing page looks off” forces the agent to guess which button, on which page, at which screen width. It may guess right, but you pay for the wrong guesses with extra rounds.
What helps is information that bridges the two views: the element the reviewer clicked, its selector, a snippet of its HTML, the page URL and the viewport. These are the details a developer would look up with browser devtools, and an agent can use them directly to search a repository.
The workflow, step by step
This example uses Symu 2 and Claude Code, but the same loop applies to any combination of feedback tool and agent.
- Get the site online at a URL. A deployed preview, staging or production URL works. A tool that loads sites from its own servers cannot reach
localhost, so use a tunnel or a preview deployment for local projects. - Add the URL to the feedback tool. In Symu 2, paste it into a new project. The site opens inside Symu 2 through a proxy, with nothing added to the site itself.
- Browse and pin comments. Switch to comment mode, click the element and write what should change. Check the tablet and mobile views as well, because each comment stores the breakpoint.
- Invite others if needed. Send the share link to a client or teammate. They comment without creating an account and their comments queue up with yours.
- Let the agent read the comments. With MCP connected, ask Claude Code to fix the open comments for the project. It reads them, edits the code, and reports back.
- Review and resolve. Reload the page at the same breakpoint, confirm the fix and resolve the comment, or reopen it with a note.
Connecting the agent with MCP
The Model Context Protocol (MCP) lets an AI assistant call tools in another application. In Symu 2 you connect Claude Code with one line from the “MCP for AI” window in the panel header:
$ claude mcp add --transport http symu https://symu.co/mcp/YOUR_TOKEN
After that the agent has these tools: list_projects, list_comments, get_comment, export_open_comments, reply_comment, resolve_comment and get_screen. A good first prompt is:
Use the symu MCP server. Export the open comments for the "Acme" website, fix them one at a time in this repository, then reply on each comment with a one-line summary of the change and resolve it.
Full setup details for Claude Code, Claude Desktop and Cursor are on the MCP page.
No MCP? Two fallbacks
- Copy as AI prompt. Every comment has a button that copies its text plus the element context as a prompt you can paste into any agent.
- Markdown export. Export all open comments of a project as one markdown task list, grouped by page, and paste it into the agent or save it in the repository.
Tips for better results
- One change per comment. Agents handle small, separate tasks better than a paragraph with five requests.
- State the outcome. “Make the hero heading fit on one line on mobile” beats “hero looks bad”.
- Mention the constraint. “Keep the existing colors” prevents creative redesigns.
- Ask the agent to reply. A short note on the comment tells the reviewer what changed and saves a message.
- Resolve yourself. Let the agent propose, and the human confirm.
For examples of vague and precise comments see how to write feedback an AI agent can act on. Building with Lovable, Bolt or v0? Start with the vibe coding workflow.