Use this checklist before publishing, merging, or claiming success from any agent-assisted work.
- Has the agent observed the real artefacts before acting?
- Has it separated the user request from inferred goals?
- Has it labelled assumptions as confirmed, inferred, or unknown?
- Has it decomposed the task into the smallest safe subproblems?
- Has it compared more than one implementation path when the decision is material?
- Has it identified a plausible failure path before implementation?
- Has it named the affected files, routes, UI states, data paths, scripts, and docs?
- Has it implemented only inside the approved scope?
- Has it avoided broad rewrites when a minimal safe change is sufficient?
- Has it run commands that actually test the changed behaviour?
- Has it avoided treating one passing command as proof for unrelated behaviour?
- Has it inspected UI output when UI quality is part of the claim?
- Has it inspected keyboard and accessibility behaviour when accessibility is part of the claim?
- Has it mapped release claims to specific artefacts and verifiers?
- Has it compared implementation evidence, behavioural evidence, and documentation evidence?
- Has it recorded failed checks instead of hiding them?
- Has it converted failures into correction requirements?
- Has it disclosed what could not be verified?
- Has it avoided “should work”, “looks good”, and “done” unless the claim is proven?
- Has it produced a final status using only verified, partially verified, not verified, or blocked?
- Has it listed files changed and why?
- Has it listed files inspected but not changed when they influenced the work?
- Has it exposed the rejected path for material decisions?
- Has it included a reviewer-reproducible evidence table?
- Has it avoided marketing language after the final status?