Top-10 AI Skills by Installs: Only 3 Actually Work — and #2 Is 157 Bytes of Nothing
Install counts for AI agent skills are a poor proxy for quality: a skill can have 1.5M downloads and still be a one-line pointer to an absent dependency. Developers who grab popular skills blindly will waste time on broken setup and may expose themselves to arbitrary code execution without realizing it.
The AI Skills ecosystem, where developers package reusable agent capabilities, has a discoverability problem that install counts do not solve. A tester pulled the top 10 skills by install volume, installed every one, and ran each skill's first documented command against clean criteria: install succeeds, self-contained, and produces real output without extra payment or keys.
Three passed outright: find-skills, hyperframes, and agent-browser. The other seven either demand sibling packages that aren't installed, require a separate CLI and login, or are so thin they merely invoke another skill that never arrives. The runner-up, grill-me, has 1.5M installs and a 157-byte SKILL.md that says only "Call the Skill tool with 'grilling'". Another skill, reddit-automation, ends with the user manually pasting text because Reddit has no comment API. The #10 skill openly requires an API key; hitting it with curl returns 401.
The author offers three pre-install checks: look at file size, grep for cross-skill references, and run the first documented command. The installer's own warning, that skills run with full agent permissions, is more honest than the ranking. Six of the top 10 come from two accounts, suggesting install numbers can be gamed or cluster around a few publishers.
Install count is a dangerously misleading ranking signal in the Skills ecosystem; it reflects marketing, packaging, or name recognition, not whether the skill works.
Several 'skills' are little more than a pointer to another skill that is not bundled, which suggests the packaging system needs a proper dependency mechanism or at least an installer that resolves the full graph.
The pattern of two accounts owning six of the top 10 slots suggests that a handful of publishers can dominate the leaderboard, making the ranking look less like a meritocracy and more like a concentration of promotion.
A skill with no scripts, agent-browser, outshines a 69-script package, so complexity in skill design does not translate to reliability — minimal scope often wins.
The honest failure modes are easier to debug: a documented paywall (requires API key) is a clear gate, while a missing transitive dependency is a silent killer that wastes a user's time.
The biggest problem with many open-source Skills is the lack of clear environment constraints and pre-validation of tool calls; switching the execution host or network environment causes them to fail outright. What actually works well are practical scripts with tightly bounded scopes.
Like.