Recommendations and tips

The playbook

Opinionated defaults, learned from the worked examples:

1. Verifier-first. Write the done-condition and the evaluator before the prompt. If you can't state how a separate checker decides "done," you're not ready to loop.
2. /goal first for babysat work. If there's a verifiable finish line, run depth-first to it and stop — don't reach for an interval loop.
3. /loop as a security camera. Use the interval primitive for genuinely ongoing patrols (a queue, a regression watch), not for work that has an end.
4. Long intervals beat short ones. A 5-minute loop is almost always wrong; most "watch" jobs are fine at 30–60 minutes and cost a fraction (see the cost math in risks and cost).
5. Graduate to durable tools. Once a loop is trustworthy, move it up the durability ladder — Codex → Automations, Claude Code → scheduled agents — so it survives you logging off.
6. Contain the blast radius. One loop per worktree/sandbox; read-only unless it has earned write; a cost cap on every loop.
7. Bet on the primitives, not the wrapper. Tools change; goal-with-an- evaluator and governed-interval are the durable ideas. Design to those.

The loop contract

Every recommendation above collapses into one artifact: a filled-in loop contract. Copy it, fill the six fields, and you have a governed loop. See also the glossary for the terms used here.

Back to the guide index.