QUICK ANSWER
Record the change you want and give the Clipy link to your coding agent. When it finishes, ask it to show the result the same way: a short recording attached to the GitHub PR. You explain the task by showing it, then review what the agent did by watching it.
Use this recording to fix the issue. Show the result with Clipy proof in the GitHub PR: [link].
The sentence that should stop being acceptable
Most agent-authored pull requests end with a line like this:
I fixed the race condition in the upload handler and verified it works correctly.
Read it as a reviewer. What was verified? Against what? Was the failure reproduced first, or did the agent run something that passed and call it verification? Would you have caught the case that broke in production?
The sentence is not dishonest. It is unfalsifiable, which is worse, because it gives the reviewer a feeling of assurance without any of the substance. And it scales badly in exactly the direction everything is currently scaling: one engineer reviewing six agent pull requests before lunch cannot re-derive each claim from scratch, so the claims get trusted by default.
"CI is green" is a narrower statement than it sounds
The obvious objection is that this is what tests are for. Tests are excellent, and they answer a specific question: do the behaviours we previously wrote down still hold?
That leaves out three things that come up constantly.
Behaviour nobody encoded. An existing test can miss a bug because its assertions or environment do not reproduce the failing case. The fix usually adds one, but the surrounding behaviour it might have disturbed is still uncovered.
Anything visual. A test asserts the element exists. That assertion alone does not show the modal is reachable, the spinner resolves, the error message is legible, or the layout survives at 390 pixels wide. Browser and visual tests can cover those checks when explicitly written. These are the failures users report and the ones tests are worst at.
Whether the test tests the bug. The subtle failure of agent-written tests is not a test that fails. It is a test that passes for a reason unrelated to the reported problem, and reads convincingly while doing it.
A recording does not replace tests. It covers the gap tests are structurally bad at: showing that the thing a person complained about no longer happens.
What should you ask the agent for?
Start with the recording link and the task. If the agent needs the Clipy skill or account access, ask it to handle setup and guide you through anything it needs. You can keep explaining what you want in ordinary language.
Verify this change in the running app. Record the verification with Clipy and attach the Clipy proof link to this GitHub PR: [PR link]. Show the problem before the fix and the result after it.
The agent handles recording and sharing. Open the proof link in the PR to see what it actually checked. Ask it to include the important screens and a reload when the change should persist. If you want the video file inside the PR as well, here is how to attach a video to a GitHub issue or pull request, and the size limits that apply.
A recording shows the run that was captured, not every possible case. If the agent cannot record in its current environment or cannot update the PR, it should say so. A walkthrough assembled from screenshots should be labeled as such, not presented as continuous screen capture.
Can the agent do this in CI or on a server?
Ask for the same outcome: verify the change and attach Clipy proof to the PR. The agent uses the capture tools available where it is running. You do not need to choose recording commands or assemble the video yourself.
A server may have a browser it can drive without a visible desktop. An agent may also have an existing video or screenshots from its verification. The skill helps it choose the available path. If it lacks a usable capture tool or the right test login, it should report that gap rather than claim it recorded the feature.
Keep manual setup details in the proof reference. The task prompt can stay one sentence.
The rule that separates proof from theatre
Most of the value is in one convention.
A caption must name a literal observed value.
Compare:
- Useless: "the fix works". It restates the claim under review, in a smaller font.
- Useful: "After: row count reads 3 and the Save button is enabled".
The second is checkable against the frame in about a second. The first requires trusting the thing you were trying to verify.
Two more rules that came out of watching this go wrong:
Show the failure too. Proof that only shows the working state cannot distinguish a fix from a screenshot of an unrelated screen. The before frame is what makes the after frame mean something.
Include the state that persists. A great many "fixed" bugs are fixed until reload. A third frame after a refresh catches a whole category of optimistic-update bugs for the cost of one screenshot.
What a recording shows the reviewer
Each caption records what the agent observed, at its timestamp, for review.
That changes what the reviewer is doing: instead of deciding whether to believe a sentence, they are looking at a frame and checking whether the caption matches it. That is a much easier judgement, and it is one a human is reliably good at.
The frames are the evidence, and the captions point the reviewer to the right moment in them.
What changes after a few weeks
Three effects, in the order they tend to show up.
Review gets faster. Ten seconds of watching replaces several minutes of reading a diff to infer whether a claim is plausible.
Agents start verifying more carefully. An agent that knows it must produce a before frame has to reproduce the failure, and reproducing the failure is the step most likely to reveal that the diagnosis was wrong. This is the underrated benefit: the requirement improves the work, not just the reporting.
Bug reports get better in the other direction. Once a team is used to short recordings as evidence, they start arriving at the front of the process too, and "cannot reproduce" becomes rarer.
Start smaller than you think
Do not mandate this repository-wide on day one. Pick the surface where a missed regression is most expensive, usually checkout, auth or onboarding, and require proof only for pull requests touching it. Run it for two weeks.
If reviewers are opening the links, widen it. If nobody opens them, you have learned something real about the value at that scope, which is worth more than a policy nobody follows.
The mechanics are in the proof documentation, including exactly what each caption records.
Clipy is the agent-readable screen recorder. Agents read recordings as structured context, and produce them as evidence that a change actually works. The format is open and documented.