← ClaudeAtlas

skill-improverlisted

Validate every skill in a repository and improve existing ones: structural validation of SKILL.md files on push, description linting against Anthropic's skill-creator guidance, and LLM-driven description optimization from trigger-eval results. Use when adding or editing skills, when skill CI fails, when a skill undertriggers or fires on the wrong queries, or when the user says skill-improver, validate skills, lint my skills, improve this skill, or optimize skill triggering. Not for creating a brand-new skill from scratch with no existing draft.
karozi/awesome-product-management-skills · ★ 2 · AI & Automation · score 56
Install: claude install-skill karozi/awesome-product-management-skills
# Skill Improver Skills decay silently: descriptions stop triggering, bodies bloat, references rot. This skill validates every skill in the repository on push and improves weak ones — forked from Anthropic's official skill-creator (anthropics/skills, MIT), reduced to the two parts that matter for an existing skill collection: validation and description optimization. ## Modes ### Validate mode (default — CI run or user asks for a skill check) 1. Run `scripts/validate_skills.py --repo-root .` for the whole repository, or `--skill <dir>` for one skill. 2. Report errors and advisories as written. Errors block; advisories are advisory unless `--strict`. 3. Fix every error directly (structural problems are mechanical: name/folder mismatch, missing frontmatter keys, dead references, oversize bodies). For advisories, apply the guidelines in `references/description-guidelines.md` and rewrite the description. 4. Re-run until clean. A skill that cannot pass without changing its behavior gets reported, not silently patched. ### Improve mode (a skill triggers wrongly — undertriggers or false-fires) 1. Assemble trigger-eval evidence first. Either run skill-creator's `run_eval.py` (upstream, anthropics/skills) or generate results in the schema defined in `references/eval-schema.md` by testing the skill's description against realistic queries: 5-10 queries that should trigger, 5 that should not. 2. Read `references/description-guidelines.md` before touching the