skill-self-review

$ARGUMENTS

Read .claude/skills/<skill-name>/SKILL.md to compare intent against what actually happened in this run.

Review — what to look for

Reflect on the just-completed run:

• Ambiguity: Did any instruction require re-reading, re-prompting, or guessing?
• Gaps: Did the skill omit a step the run needed to add on the fly?
• Friction: Did the skill ask for something that wasn't needed, or in the wrong order?
• Output quality: Was anything in the output weaker than it should have been given the skill's stated goal?
• Parallel opportunities: Were steps run sequentially that could have been parallel?

Generate 1–3 recommendations. Each must be concrete (what exact text changes), evidence-based (tied to something that happened in this run), and not already present in the skill.

---

Interactive mode (no --agentic)

Present each recommendation:

Skill: <name>
Recommendation: <what to change>
Why: <what happened this run that motivates it>
Change: <exact replacement or addition>

Ask once: "Implement any of these? [1 / 2 / 3 / all / none]"

Apply approved changes directly to the skill file. Confirm what was changed.

---

Agentic mode (--agentic)

Append to .claude/logs/skill-improvements.md — create the file if it doesn't exist:

### <YYYY-MM-DD> · <skill-name>

**Recommendation:** <what to change>
**Why:** <evidence from this run>
**Proposed change:** <exact text>

---

Do not prompt. Do not modify any skill files. One entry per recommendation.