darwin-skill
by alchaincyfdarwin-skill helps improve SKILL.md files with a repeatable loop: evaluate, revise, test, then keep or revert changes. Built for Skill Authoring, it combines rubric scoring with prompt-based validation and supports visual result outputs from repo templates and assets.
This skill scores 78/100, which means it is a solid listing candidate for directory users who want a purpose-built workflow for evaluating and improving SKILL.md files. The repository shows a real, multi-step optimization loop with clear trigger terms, test prompts, and git-based keep/revert logic, though it still leaves some adoption details implicit.
- Frontmatter gives explicit trigger terms and use cases for skill optimization tasks, making it easy for an agent to fire correctly.
- The SKILL.md describes a concrete workflow: evaluate, improve, test, human confirm, then keep or revert using git version control.
- Repository evidence includes scripts, templates, and generated visual assets, suggesting the skill is backed by an operational workflow rather than a placeholder.
- No install command is present in SKILL.md, so users may need to infer setup/usage from the README instead of the skill file itself.
- The repository is framed as experimental/test-like, so adopters should expect an optimization system rather than a narrowly scoped task skill.
Overview of darwin-skill skill
What darwin-skill does
darwin-skill is a skill for improving other SKILL.md files with a repeatable loop: evaluate structure, test effectiveness, apply changes, then keep or revert based on results. It is designed for Skill Authoring work where a plain prompt is not enough and you need a more disciplined way to raise quality.
Who should install it
Install the darwin-skill skill if you maintain multiple skills, review skills for an agent platform, or keep seeing SKILL.md files that look fine but underperform in practice. It is a good fit when your goal is not just “rewrite this” but “make this skill measurably better.”
Why it is different
The main differentiator is that darwin-skill combines static rubric scoring with real prompt-based validation. That matters if you care about output quality, not just formatting. It also uses a ratchet-style workflow, so weak edits are easier to roll back instead of being mixed into the next iteration.
How to Use darwin-skill skill
darwin-skill install and first check
Install with npx skills add alchaincyf/darwin-skill --skill darwin-skill. After installation, open SKILL.md first, then confirm the supporting docs and assets the repo actually uses: README.md, README_EN.md, docs/index.html, scripts/screenshot.mjs, and any files under templates/ and assets/.
Give it a complete skill brief
The darwin-skill usage pattern works best when you provide the target skill, the problem, and the success bar. Strong input looks like: “Optimize my SKILL.md for clearer steps, stronger frontmatter, and better test coverage; keep it compatible with Claude Code and preserve existing behavior.” Weak input like “make this better” leaves too much to guesswork.
Use a workflow, not a one-shot prompt
A practical darwin-skill guide is: identify the target skill, define the observed failure mode, run the evaluation loop, inspect the changed SKILL.md, then confirm whether the output actually improved on your test prompts. If the result regresses, revert before iterating again. That is the part that makes darwin-skill for Skill Authoring useful: it treats skill quality as something you can test, not just describe.
Read the repo in this order
Start with SKILL.md to understand the optimization rules, then read README_EN.md for the clearest positioning, then inspect templates/result-card.html and assets/chart-rubric.html to understand what the tool produces. If you want to adapt the system, check scripts/screenshot.mjs last so you know how visual outputs are generated.
darwin-skill skill FAQ
Is darwin-skill only for skill authors?
No. It is for anyone who needs to review or improve a skill with more rigor than a generic prompt. Skill authors get the most value, but reviewers and maintainers can use it to standardize quality checks.
How is it different from a normal prompt?
A normal prompt can rewrite text, but darwin-skill is built around evaluation, testing, and rollback. That makes it better when you need a repeatable darwin-skill usage loop and want to avoid “looks improved” edits that do not change results.
Is it beginner friendly?
Yes, if you can identify one skill file and describe what is going wrong. You do not need deep repo knowledge to start, but you do need a concrete target and a test prompt that reflects real use.
When should I not use it?
Do not use darwin-skill if you only need a quick wording polish, or if you cannot supply a meaningful test case. The workflow is strongest when there is a real before/after to compare.
How to Improve darwin-skill skill
Start with the biggest quality gap
The fastest way to improve darwin-skill results is to name the main weakness up front: unclear workflow, missing boundaries, weak triggers, or poor test behavior. That helps the skill focus on the part of the SKILL.md that actually limits performance.
Give better inputs, not just more text
A strong upgrade request includes the current file, the intended user, the tool environment, and one or two failing examples. For example: “This skill is for Claude Code, it fails when users ask for multi-step tasks, and the current frontmatter does not say when to use it.” That is much better than pasting a long complaint.
Watch for common failure modes
The most common mistake is asking for broad “improvement” without constraints, which can produce a prettier file that is less executable. Another failure mode is skipping test prompts, which removes the main signal that darwin-skill uses to judge whether the change was real.
Iterate with a narrow second pass
After the first output, review only one dimension at a time: trigger clarity, step ordering, boundaries, or validation quality. If the skill is close but not ready, ask for a second pass that preserves the working parts and fixes only the weak section. That is usually better than regenerating everything.