← All posts
Jan Pruszynski

Most Agent Skills don't load. We wrote three that do.

There are hundreds of thousands of free Agent Skills now, and a lot of them quietly fail — bad frontmatter, oversized bodies, or worse. Here's what separates a skill that works from one that just sits there, and three we open-sourced to prove it.

An Agent Skill is a folder with a SKILL.md file that teaches an AI agent how to do one task well. The format is a genuinely good idea — write a workflow down once, and it works across Claude Code, Codex, Cursor, Copilot, Gemini CLI, and dozens of other agents. So people wrote a lot of them. Vercel's skills.sh directory alone lists hundreds of thousands.

The catch is that a skill is just plaintext, which means it's as easy to publish a broken one as a working one. Snyk scanned around 4,000 public skills and found 36% had at least one security flaw. A big share of the rest fail for a duller reason: they don't load at all.

The ways a skill quietly fails

A skill doesn't error loudly when it's wrong. It just never activates, or loads and does nothing useful. The usual causes are boring and repeatable:

  • The name doesn't match the folder, or has an uppercase letter or an underscore. The spec allows lowercase letters, numbers, and single hyphens only. Break that and the agent skips it.
  • The description is vague. At startup an agent sees only each skill's name and description — nothing else. "Helps with PDFs" never wins against a real task. The description has to say what the skill does and when to reach for it, in the words a user would actually type.
  • The body is a wall. The whole SKILL.md loads into context on activation, so a 2,000-line skill is a tax on every run. Good skills stay under ~500 lines and push detail into referenced files that load only when needed.

None of this is hard. It's just unglamorous, and nobody checks it — so a fifth of the corpus doesn't work.

What we did about it

We build and run a fleet of agents on our own systems, so we already had skills we trusted from daily use. We cleaned up three, made them MIT, and — the part that matters — shipped the checker too.

  • skill-author teaches an agent to write a spec-valid SKILL.md, and it bundles a dependency-free validate.py. Point it at any skill, ours or yours: python3 validate.py path/to/skill. It catches exactly the frontmatter and sizing mistakes above.
  • commit-hygiene turns a messy working tree into clean, one-concern-per-commit history with messages that explain why, not just what.
  • repro-first-debugging makes the agent reproduce a failure before touching code, then prove the fix against that same repro — the discipline that stops "try this and see" loops.

All three pass their own validator. That's the whole point: a skill you can check is a skill you can trust.

# Claude Code
/plugin marketplace add xentropyai/skills

# any agent
npx skills add xentropyai/skills

The repo is github.com/xentropyai/skills. It's free and it stays free.

Where this goes

Free skills are a commodity — the text is trivially copyable, and the standard is open. What isn't a commodity is a skill that stays correct: re-tested as new models ship, with published evals and a changelog, so it doesn't silently rot the way a static file does. That's the line we're building next, for the higher-stakes workflows where "probably fine" isn't good enough.

If a tested skill that stays current is worth more to you than a free one you maintain yourself, join the waitlist. Either way, the three above are yours.