davila7__claude-code-templates/cli-tool/components/loops/engineering/skill-improvement-loop.md

name, description, category, interval, stop-condition, components, tags

name	description	category	interval	stop-condition	components	tags
skill-improvement-loop	Continuously grades the quality of your skills against a rubric and opens a focused pull request to improve the weakest one on every pass.	engineering	7d	Every skill scores at or above the quality threshold and the diffs produced by a pass become cosmetic (no meaningful improvement remains).	agent:expert-advisors/research-technical-spike agent:expert-advisors/architect-review command:git-workflow/create-pr	skills quality automation ci evaluation loop self-improvement

Skill Improvement Loop

Loop Engineering — an outer optimization loop for your skills. Instead of hand-editing skills one by one, you let an observer grade them, pick the weakest, open a small diff to improve it, and repeat until the diffs stop mattering. This is the marketplace adaptation of the "observer skill" pattern: run the inner skill, record its weaknesses, make a diff to improve it, repeat.

🎯 Goal

Keep a library of skills consistently high quality as it grows. Each pass surfaces the single weakest skill and proposes a focused, reviewable improvement — rather than a giant batch nobody reviews.

⏱️ Schedule

Suggested interval: 7d (weekly), or run manually after importing a batch of new skills.

▶️ Run it

/loop 1w "Grade every skill against the quality rubric, pick the lowest-scoring one, research best practices for it, apply a focused improvement, validate it, and open a draft PR. Continue until all skills clear the quality bar and further diffs would be cosmetic."

🔁 Iteration steps

1. Perceive (cheap triage) — score every skill with a deterministic rubric (frontmatter completeness, description clarity and trigger cues, presence of examples, body depth) and a security scan. This is token-free and ranks the whole catalog.
2. Select — pick the lowest-scoring skill (highest improvement priority) that is below the quality threshold.
3. Reason (real grade) — run component-researcher to apply the full LLM rubric and produce a prioritized improvement report with sources.
4. Act — run component-improver to apply the researched changes, preserving the skill's existing value.
5. Observe — component-reviewer and a security scan validate the change; open a focused draft PR titled improve: <skill-name>.
6. Repeat — on the next pass the triage re-ranks and the loop moves to the next weakest skill.

🛑 Stopping condition

Every skill scores at or above the quality threshold, and the diffs a pass would produce are cosmetic. Bake an exit criterion into the loop (quality bar + a cap on PRs per pass) so it does not burn tokens optimizing toward a local maximum forever.

🧩 Referenced components

• agent:expert-advisors/research-technical-spike — investigates best practices and grades the skill (the observer).
• agent:expert-advisors/architect-review — reviews the proposed change before it ships.
• command:git-workflow/create-pr — opens the focused, reviewable draft PR.

⚙️ Automating it with GitHub Actions

This repo ships a manual workflow, .github/workflows/skill-improvement-loop.yml, that runs the cheap triage in CI (heuristic rubric via scripts/skill_rubric_grade.py plus SkillSpector) and hands the shortlist to a Claude managed agent through a GitHub issue. The managed agent then drives the research → improve → review → PR steps. No API key lives in the repo — the managed agent does the LLM work. Start with create_issue: false to inspect the triage report, then flip it on once the shortlist looks right.

Note on this repo's CI path: internally this project drives those steps with repo-only agents (component-researcher, component-improver, component-reviewer) that live in .claude/agents/. They are maintainer tooling, not installable catalog components, so they are intentionally not listed in components: above. When you install this loop elsewhere, the catalog agents in components: fill those roles.

💡 Example

The weekly pass grades 870 skills, finds one with a vague description and no usage examples, researches how similar skills phrase their triggers, rewrites the description and adds a worked example, and opens improve: humanizer as a draft PR for review. Next week it moves on to the next weakest skill.