# Directive checklist

Run these 10 checks before your directive touches live traffic.

A directive that fails one of these will fail in production, usually within the first 50 inputs. Check every box, honestly, before you wire the agent up.

- [ ] 1. Categories are named and closed. The directive lists every allowed output and the agent cannot invent a new one.
- [ ] 2. Output format is exact and machine-parseable. The downstream switch can read every possible response without cleanup.
- [ ] 3. A tie-break rule exists. When an input fits 2 categories, a priority order settles it. The model never has to guess which one you meant.
- [ ] 4. An auto-reply rule exists. Out-of-office messages and automated confirmations have a named home so they stop polluting needs_reply.
- [ ] 5. An if-unsure escape hatch exists. The model has a safe default, and the safe default is the category a human reviews.
- [ ] 6. There are 2 or 3 examples, and at least 1 is a hard case with the right answer and the reason.
- [ ] 7. You tested it on 10 real inputs. Real ones, pulled from your actual inbox or sheet. Invented test cases flatter the directive.
- [ ] 8. You scored the 10 honestly. Count a near miss as a miss. 8 out of 10 or better means proceed. Below that, fix before wiring.
- [ ] 9. Every miss got diagnosed as a missing rule, not a model failure. For each wrong answer you wrote the rule that would have prevented it.
- [ ] 10. The directive is saved as a file where you can edit it. Not buried in a workflow node. When behavior needs to change, you change text, not code.

---

From Agents That Work, lesson 3. advizracademy.com