REVIEW, VERIFY AND HAND OFF
Make your agent prove the fix actually works
An agent saying a change works is a claim, not a check. The cheap version of verification is a recording of the thing being exercised, with the agent's own notes about what it observed at each point. You spend thirty seconds watching instead of ten minutes reproducing.
Quick answer
To get proof that an AI agent's fix works, ask the agent to record its verification run through Clipy, add marks with the literal values it observes, and send back the watch URL and the .arec context URL. You watch about thirty seconds instead of reproducing the fix yourself, and a reviewer's agent can read the same evidence.
Product walkthrough. Tap for sound.
What to record
This one is recorded by the agent, not by you.
It records the verification run against the real build.
It adds marks at the points that matter, each naming the value it observed.
It stops, uploads, and gives you back the watch link.
Say this while you record
Ask for literal observed values in the marks. "Looks correct" is not evidence. "Badge now reads 3" is.
The prompt to paste
Replace the link with your own recording's share URL, then paste this into Claude Code, Cursor, Codex, or whichever agent you connected.
Once the change is working, verify it and send me proof through Clipy. Record the verification run with whatever recording capability you already have, add marks with the literal values you observe at each step, and give me back the watch URL and the .arec context URL. Do not install another browser tool just for this. Do not capture anything outside the window you are verifying. Tell me in one line what the evidence proves.
This prompt goes at the end of an ordinary coding task. It does not authorize deploying anything.
What the agent gets from that link
The verification run itself
A watchable recording of the change being exercised against the real build, not a description of it.
Marks with observed values
The agent's own timestamped notes about what it saw at each step. These are driver-attested: they are what the agent reported observing, not something Clipy independently verified.
The agent context URL
The same recording as agent context a reviewer's agent can read, so it can check the evidence without watching the video.
One link for both audiences
You watch it. A reviewer watches it. Their agent reads it. Nobody reproduces anything.
What changes
| Writing it down | Recording it instead | |
|---|---|---|
| "It works now" | You reproduce it yourself to find out. | You watch thirty seconds. |
| Evidence in the PR | A screenshot, at best. | A link with the run, the marks and the agent context. |
| A reviewer's agent | Nothing to read. | Reads the same evidence you watched. |
| Regressions found later | Nobody can tell whether it ever worked. | The recording says what it did on that day. |
Tips
Connect your agent once
Run the setup at /agent-setup once. After that every Clipy share link works in the same session you are already in.
Check the frames behind each mark
A mark records what the agent said it observed, and the frames show it. Watch the recording alongside the marks, especially the first few times.
Verification is not permission to deploy
Recording evidence that a change works is a step before review, not a substitute for it, and it does not authorize shipping anything.
Scope the capture
The agent should record the window it is verifying, not the whole display. An unscoped capture picks up whatever else is on screen.
Common questions
Does the agent need a special browser tool to do this?
No. The point of the proof workflow is that it uses whatever recording or screenshot capability the agent already has. If it can produce a video file or a few screenshots, it can produce proof.
What if the agent has only screenshots?
Then it sends the screenshots with captions, which become timestamped narration on the resulting recording. It is less rich than a video and still much better than a claim.
Is the recording private?
It follows the account's normal settings. If you need to send it to a reviewer, use restricted access and invite them, because a private recording is readable by you alone and the link will not let them in. Private is the right setting when only your own agent has to read it.
Can I ask for this on every task?
Yes, and adding one sentence to your standing instructions is the usual way people do it.
Other jobs
The same job, elsewhere on Clipy
This page is the job. These cover the same ground for a specific agent, a specific team, or as a tool you can open right now.
The full list is on the use-case index. If your agent is not connected yet, start at agent setup.