Agentic PR workflow: Every pull request checked and proven before anyone reviews it

Problem

Every PR needs checks, proof that it works, and an eight-part summary for AI code review. Done by hand, those are the steps people skip. Our goal was a workflow where the coding agent does them every time.

Design challenge: checking enough without slowing work down. Every push gets the fast checks. When the PR is ready, the agent proves the change works and adds the proof to the summary.

What I did

• Made it the only way in. Developers ask the agent to push, which runs the /push skill. The repository’s rules require it, and a hook blocks opening a PR, or pushing to one, until it passes for the current commit. A PR opened another way still has to pass the required checks before it can merge.
• Asked for evidence, not “done.” For a page, the agent builds it, opens it, and screenshots it in light and dark mode, then checks it against the ticket and the design rules. An endpoint gets called, failure cases included, with each request and response pasted in.
• Added a fresh pair of eyes. A second agent that didn’t write the code reads only the diff, the ticket, and the rules. It checks whether the change does what the ticket asked and whether anything was left behind.
• Wrote the summary from the template. The skill fills in the PR template, refreshes it on every push, and never opens a second PR for the same branch. Rules changes always go in a PR of their own.

On one recent PR, the dark-mode screenshot would have shown a chart at 1.03:1 contrast. The second review would have caught a stale comment and duplicated logic. Each would have reached the reviewer already fixed.

Outcome

The pilot counts how many agent findings get fixed rather than rejected, from each PR’s findings table.

Who did what: I designed and built the skill. Everyone who opens or updates a PR in the pilot repositories uses it. The skill itself stays private; this page describes what it does.