What the source did
- Wrote a personal operating contract that syncs across machines and loads into Claude Code and Codex automatically.
- Rules: Linear ticket before every task; Claude plans, Codex reviews; isolated git worktrees; test-first; independent reviews before PRs; feature flags for production; model choice by task complexity.
- Goal: stop re-explaining how you work at the start of every session.
“stop repeating instructions and build a system that gets better with every task”— aditya (@adxtyahq), X
Summary written by Agent Cookbook. See the original for the full context.
Copy the prompt
Ours Our version, based on what aditya described. Not their wording. Fill in the [BRACKETS].
You are bound by my personal operating contract. Load it at the start of every session and follow it without me restating it.
1. Ticket first: refuse to start implementation until there is a [Linear/Jira/GitHub Issues] ticket ID and a one-line goal in the ticket.
2. Roles: Claude Code owns planning and architecture write-ups. Codex owns validation, gap-finding, test suites, and pre-PR review. Neither merges.
3. Isolation: every agent task gets its own git worktree/branch. No shared dirty tree.
4. Test-first: write or update failing tests before production code for the ticket's exit gate.
5. Review gate: before opening a PR, run an independent review (prefer the other agent) covering bugs, missing tests, and scope creep. Attach the review notes to the PR description.
6. Production: ship behind a feature flag when the change is user-facing. I flip the flag.
7. Model choice: use the stronger model for planning and hard bugs; cheaper models for bounded implementation. Document why if you promote a task to the stronger model.
8. Never send, buy, post, delete, or force-push without my explicit yes for that action.
At session start, restating: ticket ID, role you are playing today, worktree path, and the exit gate.
Approval rule
🛡
You merge PRs and flip feature flags. Sending, buying, posting, and deleting always need a human.
Set it up
- Save the contract in a file both tools read (for example `OPERATING_CONTRACT.md` plus a one-line include from `CLAUDE.md` / `AGENTS.md`).
- Sync that file to every machine you code on.
- Start one small ticket and force the full path: plan → worktree → tests → other-agent review → PR.
Common failures
⚠
- Agents skip the ticket and invent scope; keep the refuse-to-start rule.
- Reviews rubber-stamp the same agent's work; swap agents for the review gate.
Variations
- Drop Linear and use GitHub Issues only.
- Add Grok Bot as the ticket intake / status reporter, still no merge rights.
Get the 10 best agent setups. Free.
Plus new recipes every week. One email, no spam.
Unsubscribe anytime. Privacy