Agent feedback for the AI-in-the-loop workflow
Feedback for AI coding agents is its own artifact format, markdown with embedded screenshots, dictated context, and source URLs that an agent can read in one pass.
This is the canonical solutions page for agent feedback, the overarching pattern of producing product feedback in a format an AI coding agent can read and act on without human translation. If you arrived here looking for per-agent workflows (Cursor, Claude Code, Lovable, Bolt, and the rest), the /agents/ hub has a page for each. This page covers the shared artifact spec underneath all of them.
What "agent feedback" means
Agent feedback is product feedback structured so a coding agent can act on it directly. Concretely: a markdown document where each finding has a cropped screenshot referenced by a stable public URL, dictated commentary captured as editable text, and the source URL of the page being reviewed. The agent reads the markdown, fetches the images inline as visual context, and produces a patch.
The point is that the feedback is already in the input format the agent uses. There is no translation step between "I see the problem" and "the agent has the context." No transcribing a Loom. No screenshotting into Slack and retyping the description into a Linear ticket and then pasting both into Cursor. The artifact you produce during review is the artifact the agent ingests.
Why the format matters now
A year ago, product feedback was a handoff between two humans. The reviewer sent a Loom or a Figma comment, the engineer watched or read it, and the engineer wrote the code. The format only needed to be human-readable.
That handoff has changed. AI coding agents, Cursor, Claude Code, Lovable, Bolt, v0, Windsurf, now sit between the person finding the issue and the code that fixes it. Sometimes the agent does the whole fix. Sometimes a developer drives the agent. Either way, the agent reads the feedback before code gets written.
Agents are good at acting on structured, multimodal input. They are bad at acting on a 90-second video, a paragraph that mixes three findings into one wall of text, or a screenshot pasted into Slack with no caption. The format determines whether the agent acts on what you found or stops to ask clarifying questions. Asking clarifying questions costs a round trip. Multiply that by five findings and a one-hour review becomes a half-day.
The teams that figured this out early are not faster because their agent is smarter. They are faster because the feedback they hand the agent does not need to be reformatted first. The format is the product.
The shape of an agent-readable artifact is consistent across tools: markdown for the document, image URLs for the visuals, plain prose for the commentary, and a source URL anchoring each finding to a real page. Get those four right and any agent that takes markdown input can act on the result.
What an agent-feedback artifact looks like
A real CobaltCapture review for a staging build looks roughly like this when exported as markdown:
# Staging review, checkout flow
Source: https://staging.example.com/checkout
The submit button overflows its container at widths under 380px. Tested
on iOS Safari and Chrome Android. The container is `.checkout-cta` in
`CheckoutFooter.tsx`. Likely the `min-width` on the button needs to drop
or the container needs `flex-wrap`.
Email validation fires on every keystroke instead of debouncing. The
error flashes red while the user is still typing the domain. Should wait
for blur, or debounce 400ms.
On iOS the discount field gets covered by the keyboard when focused. The
page does not scroll to keep the input in view. Either scroll-into-view
on focus or move the field higher in the form.
That document is the input to the next prompt. The agent gets the visual context, the verbal context, and the location, all in one paste.
How CobaltCapture produces it
Three steps. No install.
Open the page you want to review. Open cobaltcapture.com in a new tab and hit Capture screen. Pick the window. Drag a box around the broken part. Hit Dictate and talk through the problem out loud, the browser turns your voice into editable text next to the screenshot. Repeat for each finding. Hit Publish.
You get back two things: a public URL (cobaltcapture.com/r/<slug>) and a markdown export. Either one drops directly into your agent's prompt. The URL is easier to paste; the .md file is easier to commit into a repo. Both contain the same content.
That is the whole product. Browser-native capture, browser-native dictation, public markdown URLs hosted on a CDN. The output is the input format the agent already uses.
Where to go next
If you have a specific agent in your stack, the per-agent pages walk through the prompt patterns that work best for each one. Start with feedback for Claude Code if you live in the terminal, or the agents hub for the full list, Cursor, Windsurf, Lovable, Bolt, v0, Replit Agent, Cline, Aider, and Zed each have their own page.
If you arrived from a tool comparison, the related solutions pages cover the alternatives directly: Loom alternative for video-first feedback teams, Scribe alternative for teams currently producing HTML SOPs, and screen capture to markdown for the broader category of turning what is on screen into a markdown artifact.
The fastest way to understand what this looks like in practice is to capture one review yourself. It takes about a minute and the output goes straight into your next agent prompt.
Capture your first review.
About a minute from open tab to a shareable URL your agent can ingest.
Start capturing