Unofficial checklist: turn a skill learned from one demonstration into instructions you can rely on reading. 1. Write [EXPECTED_RESULT] first: what the finished work looks like and which details you will compare. 2. Open [SKILL_NAME], the draft made from your demonstration. Read it against what you actually did. A draft from one pass knows one path only. 3. Check that no login, code or card number was typed during the recording. If one was, remove it from the draft and change that credential. Sign-ins go through the secure handoff, not the recording. 4. Name [INPUT_SOURCE] in the draft: which site or app, which account, and how to tell the right record from a similar one. 5. Add decision rules. For each place where you chose by eye, write the condition. Add when to stop because no choice is clearly right. 6. Add failure handling: which step failed, what was already done, what is left. A finished step is not repeated. 7. Add a missing-data rule: unreadable, absent or old input is reported as a gap. Nothing is filled in from memory. 8. Add the approval boundary: stop before sending, publishing, purchasing, paying, deleting or changing production. 9. Add a description. Teammates and their Bots use it to decide when the skill applies, and a skill cannot be published to a team without one. 10. Pick [SAFE_EXAMPLE]: real enough to exercise the steps, with no private data and no way to cause an outside action. 11. Run the skill once on it. Do not approve any outside action during this run. 12. Compare the output with [EXPECTED_RESULT]. Write down each difference and fix the draft, then run it again. 13. Repeat the safe run whenever the site, app or source format changes. Until it passes again, treat the skill as a draft.