skill-retrolisted

# /skill-retro One evidence-driven retro pass over this kit's skills. The core rule: **no evidence, no change** — skills are never rewritten on calendar cadence or taste; every proposed edit must cite a concrete friction, outcome, or docs drift. This skill is itself in scope (dogfooding). Respond in the user's conversation language. Skip narration between tool calls — this may run unattended as a scheduled routine. ## 1. Enumerate scope Skills = directories with a `SKILL.md` in the claude-kit checkout's `skills/`. Locate the checkout: if `~/.claude/skills` is a symlink (the kit's symlink install), follow it to the repo root; otherwise ask the user where the claude-kit checkout lives. If no writable checkout exists (plugin-only install), still run the evaluation and report the verdicts — skip only the PR delivery step. Also note the always-loaded rules skills rely on (the kit's `rules/*.md`, and the user's global CLAUDE.md sections such as Plan & Delegate / Skill Feedback Capture) — the thing that needs fixing may be a rule, not a skill. ## 2. Gather evidence (cheap first, expensive only if warranted) 1. **Friction logs** — read `~/.claude/skill-feedback/*.md`. This is the primary signal (self-reported deviations, user corrections, uncovered judgment calls). 2. **Invocation traces** — mine `~/.claude/projects/*/` session logs for the last month: count invocations per skill (search for skill names / `<command-name>` markers). For skills with hits, sample at mo