We AI-reviewed 10,000 Claude skills. Here’s what separates the good ones.

TL;DR. Our auto-review pipeline has now scored just past 10,000 Claude skills, drawn from a catalog of 72,587 indexed across five agent platforms. The headline finding isn’t that most skills are bad — it’s that most are interchangeable. Three signals separate the skills that score 85+ from the long middle: a distinctive, opinionated judgment encoded in the rubric; a small tool surface that runs on a stock install; and a reproducible verify-it-yourself loop.

On this page

1. What we actually built
2. What 10,000 reviews look like
3. What separates a high-scoring skill
4. What automated review can’t see
5. If you’re writing a skill
6. Related reading
7. FAQ

What we actually built

SkillHub indexes Claude Code skills — the SKILL.md bundles that teach an agent a repeatable capability — from across GitHub, plus the equivalents for Cursor, Windsurf, Copilot, and Codex. Indexing is cheap; judgment is not. So we built an automated review pipeline: it fetches a skill’s source, runs it through a scoring rubric on a 0–100 scale, records a verdict, and re-queues the skill for another pass whenever the upstream source changes.

The rubric is not a linter. It asks the questions a careful reviewer would: does this skill do something a plain prompt couldn’t? Is its tool surface justified, or is it a thin wrapper around one API call? Would following it actually produce the result it promises? Skills enter the review queue only above a baseline heuristic score, and the reviewer assigns the final aiScore you see on each catalog page.

What 10,000 reviews look like

Here is the honest state of the pipeline as of this writing — these figures are pulled live, and the running tally is public on the review stats page.

Two numbers in that table do more work than the rest. The gap between 13,979 runs and 10,050 skills is the cost of drift: a skill scored last month may not be the same skill today, so a quarter of our review effort goes to re-scoring things we have already seen. And the 813 currently flagged are a standing reminder that a review is a snapshot, not a certificate.

What separates a high-scoring skill

Skills do not spread evenly across the 0–100 range. They bunch. The vast middle is competent and forgettable — a skill that restates what a capable model would do anyway. The top of the catalog is narrow, and the skills that reach it share three traits.

1. Encoded judgment. The highest-scoring design skill in the catalog, garrytan/gstack/design-review (86), earns its place not by knowing CSS but by carrying a designer’s opinion the model would never volunteer on its own — an explicit detector for AI-slop patterns like dead-center compositions and default gradients. Anthropic’s own hook-development (89, the catalog’s top score) does the same for event-driven automation: it encodes which hook fires when, a distinction most agents get wrong by default.

2. A small tool surface. The skills that score well tend to run on a plain Claude Code install — Bash, file edits, search, web — and reserve heavier machinery for when it genuinely earns its keep. trailofbits/skills/codeql (87) wraps a real, heavy analysis engine, and scores well precisely because the wrapping is thin and the payoff is large. A bespoke server that saves you one prompt is a liability; one that unlocks CodeQL is not.

3. Reproducible verification. NousResearch/hermes-agent/research-paper-writing (88) and mem0ai/mem0/mem0-integrate (86) both score highly for the same unglamorous reason: they tell the agent how to check its own output, instead of declaring victory after the last step. A skill that ends at “now write the section” scores worse than one that ends at “now write the section, then verify every citation resolves.”

What automated review can’t see

An honest accounting of the limits matters more than the headline.

• Survivorship. Skills that fail badly — or turn out to be malicious — get pulled from the browsable catalog. So the scores you see skew high not because most skills are good, but because the worst ones are no longer in the room. (We learned that lesson the hard way; see our earlier advisory on malware in the openclaw/skills repo.)
• Drift. A score describes a skill at a moment. Upstream authors edit, refactor, and occasionally gut their skills. Our 813 flagged-for-re-review skills are the visible edge of a larger truth: every verdict is decaying from the moment it’s written.
• Taste. The rubric can spot a missing verification step or an unjustified server. It cannot tell you whether a skill is tasteful — whether its judgment is one you’d actually want. For that, a human still has to read it. The pipeline’s real job is to decide which ten skills out of ten thousand are worth that human read.

If you’re writing a skill

Three things, drawn directly from what scores well:

1. Encode an opinion, not a summary. If a capable model would do it anyway without your skill, your skill isn’t earning its install. Write down the judgment call the model gets wrong by default.
2. Justify every tool. Default to a stock install. Each extra dependency or bespoke server has to pay for itself in capability, not convenience.
3. Tell the agent how to check itself. End on a verification step, not an action. The single most reliable predictor of a high score in our corpus is a built-in way to prove the work is done.

Related reading

• Skill spotlight: hook-development — the catalog’s top-scoring skill (89), and a worked example of encoded judgment.
• Skill spotlight: design-review — how an AI-slop detector turns taste into a repeatable loop (86).
• Malware found in AI skills — why survivorship bias in the catalog is not academic.

Browse the live, scored catalog at skills.palebluedot.live, skim the full list of AI-reviewed skills, or read more about the method on our about page.

FAQ