Capstone: one safe, cheap task

Time to run one real task well. Don't set up ten — set up one you'd actually trust, using every habit from this course. Here's the checklist, in order; it's the whole course on one page.

1. Pick the agent by the job (Module 4): availability first, then free-vs-paid, then match the shape — clock (Grok Bot), guarded errand (Muse), continuous goal (Dots).
2. Scope the access (Module 1): connect only the apps the task needs; prefer a scoped or read-only account; on Grok Bot, remember a separate Bot is not a wall.
3. Make it cheap (Module 2): choose the slowest cadence or least-continuous goal that works; trigger on a real event where you can; add a cheap "did anything change?" check before expensive work.
4. Decide the endings (Module 3): sort every action into allow automatically (safe, reversible) or stop at a draft (risky, irreversible); set firm "ask first" limits on anything that sends, buys, publishes, or deletes.
5. Add the trust habits (Module 3): a no-data/stale-data rule, repeatable actions that don't pile up, and a plan to read the audit trail.
6. Test read-only first (Module 4): point it at a safe target, run it looking-and-proposing before acting, work the pre-flight checklist, then grant real action.
7. Watch the first runs, then let it go — and review it occasionally. If it's doing nothing useful most of the time, slow it down or retire it.

That's the discipline: pick by the job, scope it tight, make it cheap, decide its endings, make it trustworthy, test it safe, and watch it in. Do this once and you'll never set up a personal agent the careless way again — on Grok Bot, Muse, Dots, or whatever ships next.