Improve
Sub-Skills
- Trigger Conditions
- Phase 1: COLLECT — Gather Session Signals (+6)
- Decision Logic (Scoring)
- Creating New Skills (+3)
- Integration
- Scope
- Related Commands
- Manual vs Automatic Invocation
Iron Law
No improvement shall be applied without a scored candidate entry and evidence of the session signal that triggered it — no speculative edits to skills, rules, or config.
Rationalization Defense
| Excuse | Reality |
|---|---|
| "This improvement is obviously correct, no need to score it" | Obvious-seeming improvements have broken skills before. The scoring gate exists to catch false positives. |
| "I'll create the candidate entry after I apply the change" | Post-hoc justification is not a gate — it is rubber-stamping. Score first, apply second. |
| "The user asked me to improve this skill directly" | Direct user requests bypass the nightly pipeline, but they do not bypass the evidence requirement. Show the signal before editing. |
| "This is just a typo fix, not a real improvement" | Typo fixes are fine inline. But if you are changing behavior, logic, or constraints, it is an improvement and needs a candidate entry. |
Red Flags
These phrases signal you are about to violate the Iron Law:
- "while I'm here, let me also update this skill"
- "this improvement is straightforward enough to apply directly"
- "I noticed this could be better"
- "no need for a candidate entry — this is minor"