What the source did
- Created changes.log that both tools read first and update last.
- Kept Claude's rules in CLAUDE.md; Codex imported them into AGENTS.md.
- Ran a live round trip on a small calculator: Claude added a breakdown, Codex turned it into a chart and checked it in the browser, Claude then rebranded it.
“It gives the next agent something to judge.”— Wyndo and Dheeraj Sharma, Substack
Summary written by Agent Cookbook. See the original for the full context.
Copy the prompt
Ours Our version, based on what Wyndo and Dheeraj Sharma described. Not their wording. Fill in the [BRACKETS].
Before you start: read the LAST entry of changes.log and restate in one line what you're picking up.
When you finish, append an entry to changes.log:
[YYYY-MM-DD HH:MM] [AGENT NAME] [TASK ID]
Status: DONE | PARTIAL | BLOCKED
What changed and why:
Files touched:
Unfinished:
Don't undo:
Next agent should: (one concrete next step + the command that proves it works)
Also review the previous agent's work and list anything you'd push back on.
Approval rule
🛡
Neither agent merges or deploys. You approve each hand-off after reading the new log entry.
Set it up
- Pick a low-risk repo both tools can open.
- Add changes.log and paste the block above into both CLAUDE.md and AGENTS.md (or import one from the other).
- Run one small round trip: agent A builds, agent B reviews and improves, agent A picks back up.
- Read each entry before approving the next run.
Common failures
⚠
- Both agents run at once and overwrite each other's entries. Hand off sequentially or use separate branches.
- Vague entries ('fixed stuff') give the next agent nothing to judge.
- Rules drift when CLAUDE.md and AGENTS.md diverge; keep one source and import it.
Variations
- Swap direction: Codex drafts, Claude Code wires it in.
- Add a 'reviewer findings' section to every entry.
Get the 10 best agent setups. Free.
Plus new recipes every week. One email, no spam.
Unsubscribe anytime. Privacy