skill-auto-improvelisted

# Skill Auto-Improve This is not a skill you run. It's a behavior you follow every time you execute any skill. --- ## the rule After delivering output from ANY skill, silently evaluate that output against the skill's own instructions. If something failed, log it. When failures pile up, fix the skill. That's it. --- ## what counts as a failure A failure is when the output violates the skill's own rules or the user rejects/corrects the output. Specifically: 1. **Skill rule violation** — the output breaks an explicit instruction in the skill (e.g., skill says "never use semicolons" and the output has semicolons) 2. **User correction** — the user says "no, redo this," "that's wrong," "change X to Y," or otherwise rejects the output 3. **User edit-after-delivery** — the user accepts the output but immediately asks for specific changes that the skill should have handled 4. **Repeated manual instruction** — the user gives the same guidance that isn't in the skill (e.g., they keep saying "make it shorter" every time you run the email skill) **What is NOT a failure:** - User changes their mind about the brief (that's a scope change, not a skill failure) - Output is correct but the user wants a different creative direction - One-off edge cases that won't recur --- ## step 1: silent eval (every skill run) After delivering skill output, do a quick internal check — do NOT output this to the user: 1. Re-read the skill's key rules (you already have them in context) 2. Check: di