The Top 10 Agent Skills on skills.sh, Actually Tested
Agent skills are becoming the new unit of open-source distribution in the AI coding era — encapsulated expertise that turns a generalist AI into a specialist. Knowing which ones actually work, rather than which ones are popular, saves hours of trial and error and directly affects the quality of AI-generated code, designs, and architectural decisions.
Vercel's skills.sh has become the de facto app store for AI agent capabilities, with the top-ranked skill hitting 2.4 million popularity points. A developer installed and tested all Top 10 skills across real-world scenarios — from frontend design and React performance optimization to architecture review and browser automation.
The results show a landscape dominated by Vercel (4 slots) and TypeScript expert Matt Pocock (3 slots), with Anthropic and Microsoft each claiming one spot. The #1 skill, find-skills, is the CLI gateway that makes the entire ecosystem usable; the #2, Anthropic's frontend-design, specifically cures the cookie-cutter "AI look" by enforcing bold aesthetic direction before writing any code.
Practical takeaways include recommended skill combinations for different roles — frontend developers, architects, React power users, and content creators — plus a growing list of noteworthy skills beyond the Top 10. The ecosystem is maturing fast, and the best way to use it is to let your AI agent browse skills.sh and install what fits your workflow.
Popularity rankings on skills.sh conflate install base with genuine utility — microsoft-foundry ranks #7 but is useless to anyone outside the Azure ecosystem, while grill-me at #5 may be the most intellectually valuable skill in the top 10 for architects.
The dominance of Vercel and Matt Pocock in the top 10 suggests the ecosystem is still supply-constrained: a small number of prolific creators are defining what agent skills look like, rather than a broad community.
frontend-design's approach — choosing an aesthetic direction before generating code — mirrors how human designers work, suggesting that effective AI skills encode process constraints, not just knowledge.
The `file:line number + issue` output format used by web-design-guidelines is a design choice that prioritizes actionability over explanation, a pattern other developer tools could adopt.
agent-browser's session continuity across commands solves a real pain point in browser automation that traditional scripting approaches handle poorly, making it a genuine step forward for AI-driven web interaction.
Skill installation is already self-serve: an AI agent can browse skills.sh, compare options, and install skills without the user touching a command line, which changes the discovery model from 'search and evaluate' to 'delegate and verify.'