Welcome to the AI Collab Bridge. Below is your playbook.
If you don't know your role yet, look at the invocation:
- Given a diff and asked to review → you are the reviewer.
- Finished a task and want a second pair of eyes → you are the implementer.
Claude (or another AI) has handed off work to you. The packet contains a summary, optional focus questions, a file list, and the diff.
- Read the whole diff. Don't trust the summary alone.
- Apply your strengths. Codex tends to do well at:
- Build sanity: does this actually compile? Are imports right? Are types coherent?
- Test coverage: are tests present? Do they actually catch regressions, or are they tautological?
- Scripting / automation: if the change touches a script or CI, will it do what it claims under the conditions it'll actually run in?
- Repo hygiene: stray debug logs, commented-out code,
.DS_Store, accidental large binaries, broken imports. - Determinism: did the change introduce flakiness, time/clock dependencies, ordering assumptions, or hidden state?
- Respond using
templates/review-response.md. Usefile:linerefs. Every finding gets a clear reason. - Be honest. Approve when the work deserves it. The bridge only works if the verdict is meaningful.
- Don't critique architecture you don't have context on. If you suspect a design problem but lack the surrounding code, flag it as a non-blocking suggestion.
- Don't pile on findings. The response is a list of real issues, not a comprehensive checklist.
- Don't claim to have executed code you didn't actually run. Be explicit in
What I did NOT check.
You finished work and want Claude (or another AI) to look it over.
- Stage the packet:
PACKET=$(mktemp -t packet) SUMMARY="What the change does in one or two sentences" \ QUESTIONS="Anything specific you want focused on" \ <skill-root>/scripts/stage-packet.sh <base-ref> > "$PACKET"
- Request the review:
(
<skill-root>/scripts/request-review.sh auto "$PACKET"
autopicks a healthy CLI that isn't you; nameclaude/geminito choose explicitly.) - Read the response. Disagree with reasoning, not dismissal.
- Re-request after material changes. Pass the previous response via
PREV_RESPONSE=<file>so the reviewer sees what round 1 said.
- Don't ship past
BLOCKwithout resolving it or making the case for an override (politely, with reasoning). - Don't treat the reviewer's findings as inviolable. Treat them as smart input from another mind — incorporate what's right, push back on what's wrong.