AI agent feedback workflow: from client click to shipped fix
If you ship with Cursor, Claude Code or Codex, the bottleneck has moved. The agent can make the change in minutes — but someone still has to turn "the spacing looks off under the pricing cards" into a prompt with a page, a component and an element in it. This article shows a workflow where that translation step disappears.
The problem: feedback is written for humans
Screenshots with arrows, Loom videos, "third section from the top" — all of it assumes a human will hunt down the element. An agent can't click through a PDF. So you re-describe every note by hand, once per fix, forever.
The workflow
- Install once. Drop one script tag on the site — or paste the install instruction into your agent and let it do the setup itself.
- Collect notes on the live page. You, your client or your users click the element and type a sentence. Better Feedback captures the page URL, the clicked element, its CSS selector, the surrounding text, the React component stack and the viewport — automatically.
- Export the batch as markdown. One request returns every open note as a structured task list. Selectors map to the codebase; page paths map to route files.
- Let the agent work the list. Paste the export into Cursor, Claude Code or Codex — or let the agent fetch it itself over the API and resolve each item with a note as fixes land.
Why element context beats screenshots
A screenshot tells the agent what the page looked like. A captured element tells it where the change goes: the selector, the component name and the nearby text let it find the exact code without guessing. Notes anchor to the DOM element, so they survive redeploys, animations and layout shifts — a coordinate-based pin doesn't.
Keeping the agent's context small
Better Feedback is deliberately curl-first: the agent fetches a short playbook on demand instead of carrying a heavyweight tool definition in every session. One line in your project notes is all it needs, and the workspace key stays in your shell — never in the repo.
The full loop — client clicks, agent ships — is what Better Feedback for AI coding is built around. Three projects are free, so you can wire it into one repo tonight.