• A 12-minute proof protocol for AI-written pull requests
  • 1. Name one real user path
  • 2. Start clean
  • 3. Attack the seams
  • 4. Save evidence
  • 5. Reset after every fix

§A 12-minute proof protocol for AI-written pull requests

AI coding agents are often strong at local transformations and weak at proving that the changed system works as a whole. A green unit test can coexist with a broken route, stale process, hidden redirect, invalid mobile layout, or retry bug. The useful question is not “did the agent write plausible code?” It is “what evidence would make this change safe to ship?”

§1. Name one real user path

Write the shortest valuable journey in observable terms: start the service, open the public URL, perform the action, and verify the durable result. Test that governing path first. If it fails, downstream polish does not matter yet.

§2. Start clean

Stop only the disposable process created by the test, clear only isolated test state, and launch from the same entry point a real operator uses. A manual resume or inherited process can hide boot failures. Log the exact process, port, executable, and state directory so the run is auditable.

§3. Attack the seams

Most agent regressions live between components. Check ownership, retries, duplicate submissions, concurrency, stale caches, redirects, authentication boundaries, and persistence across a restart. Use at least two different test styles because a single harness can share the product’s blind spot.

§4. Save evidence

Keep request results, relevant server logs, and headless desktop and mobile screenshots. Record which source URL was used and whether the final URL changed. Evidence should let another person distinguish a real pass from a mocked or partially resumed run.

§5. Reset after every fix

If a bug is found after a live run, that run no longer proves the final build. Fix the root cause, then restart the count at zero and repeat the entire journey from a fresh start.

The free Agent PR Proof Pack packages this protocol as two copy-ready prompts and a small release-evidence record. It unlocks after a GitHub-verified email opt-in; maximum one email per month and unsubscribe any time: https://tensola.com/playbooks/agent-pr-proof-pack?source=nostr-article