skill-from-feedbacklisted

# Skill From Feedback Improve an **existing** skill using concrete feedback. The skill file is the product under test; feedback is the bug report / UX note / eval miss. ## Safety - Identify the **canonical source** skill (repo `src/skills/<dir>/` or the user’s source of truth). Do not “fix” only a generated install copy under `~/.claude/skills` / `~/.codex/skills` / `~/.grok/skills` without updating the source and re-syncing. - Preserve the skill’s public `name` (kebab-case id) unless the user explicitly renames and updates the catalog/install surface. - Keep frontmatter valid: `description` ≤ 1024 chars; flow skills start with `Use when `; body has Safety / Workflow / Verification when `type: flow`. - Do not invent user requirements. If feedback is vague, ask one focused clarifying question or propose 2–3 concrete patches and wait for pick. - Do not delete safety gates, remote-delete bans, or main-branch protections to “make it easier.” - After edits, run the repo’s skill validation (`skills.test.ts` / catalog checks) when this is the dufflebag monorepo. ## Workflow 1. **Capture feedback** in one place: user quote, failing session behavior, missing trigger phrase, wrong handoff, over-trigger, or eval miss. Note agent host (Codex / Grok / Claude) if relevant. 2. **Locate the skill**: open `SKILL.md` (+ references/scripts). Summarize what it claims today (triggers, steps, done criteria). 3. **Diagnose class of gap** (pick one primary): - **Discovery** — description/tri