Capstone: build one safe, cheap, quiet routine

Time to put it all together. You're going to design one real routine using every principle from this course. Don't build ten; build one you'd actually trust, and build it right. Here's the full checklist, in order — it's the course on one page.

1. Prove it by hand first. Do the task once as a normal chat. Make it reliable. Only when it works cleanly, save it as a skill (the how) — then automate it. Never schedule a task you've never completed successfully.

2. Choose the trigger. Ask the decision question: can the source send an event? If yes, write a narrow trigger (match one specific thing, one action, one hard stop). If no, use a schedule — and pick the slowest cadence the job can live with, not five minutes "just in case."

3. If it's fast, add a change check. For any schedule faster than hourly, put a cheap check in front: read one small signal, compare to last time's saved value, stop if unchanged ("do not take a quick look around"), and only do the full run when something really changed.

4. Decide the ending in advance. List every action. For each: is it safe and reversible, or risky and one-way? Safe ones get an Allow automatically rule so the routine finishes unattended; risky ones stop at a draft for you to approve. Add firm Ask first rules on anything that sends, publishes, buys, deletes, or touches production — they win over any allow rule.

5. Make it trustworthy when things go wrong. Add a no-data / stale-data policy ("report the failure instead of using old data"), make the actions idempotent (safe to run twice — no duplicate posts), and tell it where to report partial completion.

6. Keep it to least privilege. Connect only the tools this routine needs; prefer a scoped account over your full-access one; remember every Bot shares one computer, so don't rely on a separate Bot for isolation.

7. Add the run log. Append one line per run to routines/log/<name>.csv so you have history beyond Grok Bot's 20-run window.

8. Test on a safe copy, then enable. Remember a test run does real work — point it at a safe target, run the review checklist (current inputs, right format, audit trail, stops at approval, explicit failures), fix anything that's "no," and only then turn it on.

9. Review it weekly. Watch the log. If it's almost always "no change," slow it down or kill it. A small set of trusted routines beats a pile of forgotten ones.

That's the whole discipline: prove it, slow it, check before it works, decide its 3 a.m. ending, scope it, log it, test it, and prune it. Do this for one routine and you'll never set up another the careless way.