Jason-hub-star Goal Contract
Create one brief that Codex, Claude Code, or another agent can execute without changing the finish line.
Required contract
- Outcome — measurable state that must become true.
- Verification — runnable commands or inspectable artifacts that prove it.
- Constraints — behavior or quality that must not regress.
- Boundaries — allowed and forbidden files, systems, data, and actions.
- Iteration policy — how failed passes choose the next smallest change.
- Blocked stop — when the agent must stop and report instead of guessing.
Human taste or approval belongs between goals, not inside an automated completion check.
Output
Write a compact one-line goal plus a Markdown brief containing the six fields and an execution log. Use shell-runnable verification when possible; do not make an internal agent tool the only proof of success.
Next
Use phase-loop when the goal becomes three or more ordered implementation phases. If verification cannot be written, return to converge-plan.