The SKILL.md ecosystem just hit escape velocity: skills.sh now tracks 1,436,077 installs across the open directory, OpenAI has quietly started shipping official SKILL.md skills, and enterprise suites (Feishu/Lark at 14.9M installs, Microsoft Azure at 8.0M) are proving skills are no longer a Claude Code curiosity. Also this week: Snyk's ToxicSkills audit keeps casting a shadow โ 1,467 malicious skills flagged, with community analysis putting "some form of dangerous issue" at ~36% of audited skills. Audit before you install. Now, the podium.
| Rank | Skill | Developer | Installs | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | vercel-react-best-practices | Vercel Labs | 703,900 | SKILL.md (vercel-labs/agent-skills) | The React conscience your agent lacks |
| ๐ฅ | domain-modeling | Matt Pocock | 609,600 | SKILL.md (mattpocock/skills) | Entities before code, always |
| ๐ฅ | diagnosing-bugs | Matt Pocock | 580,200 | SKILL.md (mattpocock/skills) | Root-cause discipline on demand |
Why these ranked here: Vercel's React best-practices skill is the highest-ranked coding skill on skills.sh not currently in a review cooldown, and it earns it โ it encodes the exact conventions that stop coding agents from their worst over-engineering instincts. Pocock takes two podium spots because his per-skill install counts are staggering: domain-modeling (609.6K) forces architectural thinking before implementation, while diagnosing-bugs (580.2K) imposes a systematic hypothesis-test loop instead of random patching. Note: defending champion ponytail (Dietrich Gebert) rotates off this week per the 30-day rule and returns October 3.
| Rank | Skill | Developer | Installs | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | frontend-design | Anthropic | 874,400 | SKILL.md (anthropics/skills, ~149K โญ) | The official taste layer |
| ๐ฅ | web-design-guidelines | Vercel Labs | 623,800 | SKILL.md (vercel-labs/agent-skills) | Guardrails for generated UI |
| ๐ฅ | design-md | Google Labs (google-labs-code) | New entry | SKILL.md | DESIGN.md files go official |
Why these ranked here: Anthropic's frontend-design is the #1 design skill on skills.sh (874.4K installs) and the strongest single skill in the official repo โ it's the difference between a generated UI that looks like a 2019 admin panel and one that ships. Vercel's web-design-guidelines (623.8K) pairs with it perfectly as a rules layer. The surprise bronze: Google Labs has entered the ring with design-md, a skill for creating and managing DESIGN.md files โ evidence that every major lab now treats SKILL.md as table stakes.
| Rank | Skill | Developer | Stars/Installs | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | Higgsfield Skills | Higgsfield (higgsfield-ai) | New entry | SKILL.md suite (higgsfield-ai/skills) | 30+ media models in one skill pack |
| ๐ฅ | youtube-clipper-skill | op7418 | Trending | SKILL.md + scripts | Semantic video clipping that actually works |
| ๐ฅ | speech | OpenAI | New entry | SKILL.md (openai/skills) | The format war is over |
Why these ranked here: Higgsfield Skills is the week's biggest media launch โ one skill pack giving any agent 30+ image/video models, character training, 4K stills, and a "Virality Predictor" that scores video before you post. youtube-clipper-skill (op7418) does something genuinely hard: semantic highlight extraction from long-form video with bilingual subtitle embedding via yt-dlp + FFmpeg โ not dumb timestamp cutting. And bronze goes to a milestone more than a metric: OpenAI now publishes SKILL.md skills, with speech (TTS/narration) in its official repo since late August. When OpenAI adopts your packaging format, the format has won.
| Rank | Skill | Developer | Platform | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | notebooklm-skill | PleasePrompto | GitHub | SKILL.md + Patchright/Playwright | Your knowledge base, finally wired into the agent |
| ๐ฅ | fact-checker | davila7 (claude-code-templates) | GitHub | Agent template/SKILL.md | Verification as a first-class citizen |
| ๐ฅ | ToolUniverse | Zitnik Lab, Harvard Medical School | Self-hosted docs | Skills collection | Expert-level science workflows |
Why these ranked here: notebooklm-skill is this week's most-clipped research tool โ it connects Claude Code directly to Google NotebookLM through persistent browser automation, killing the copy-paste-between-browser-and-terminal loop entirely. fact-checker from davila7's deep-research-team gives any agent claim-verification, source-validation, and misinformation-detection behavior out of the box โ exactly what the ToxicSkills era demands. ToolUniverse from Harvard's Zitnik Lab packages sophisticated scientific research workflows (multi-tool, expert-level) as agent skills โ academia is shipping SKILL.md now.
| Rank | Skill | Developer | Platform | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | Wondel.ai Skills | Wondel.ai | GitHub + skills.wondel.ai | 50 skills + 12 guided journeys | Bestselling books as installable method |
| ๐ฅ | TWG CLI agent skills | Atlassian | Official docs + atlassian/twg-cli | SKILL.md via twg skills install |
Jira/Confluence/Bitbucket in one prompt |
| ๐ฅ | skill-from-masters | GBSOSS | GitHub | Meta-skill (SKILL.md) | A skill that writes skills |
Why these ranked here: Wondel.ai packages 51 frameworks from bestselling business and engineering books as skills, plus 12 "guided journeys" that orchestrate skills step-by-step โ it's trending hard this week and represents the "knowledge-as-capability" wave. Atlassian's TWG CLI skills are the sleeper enterprise story: official Jira, Confluence, and Bitbucket skills you install with one command, with docs refreshed this week (open beta since May). Bronze to skill-from-masters, a meta-skill that researches domain experts, extracts their methodologies, and generates a professional-grade SKILL.md โ recursive capability creation.
| Rank | Skill | Developer | Platform | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | rtk | Open source (JetBrains-benchmarked) | Blog benchmark + repo | Skill/tool wrapper | 60โ90% token savings, measured |
| ๐ฅ | MD Compressor | Community | Viral tutorial | SKILL.md | Fit more skills in less context |
| ๐ฅ | context-compression-strategies | Community (MCP Market listing) | Claude Code skill | SKILL.md | Long-session context summarization |
Why these ranked here: rtk posted the hardest numbers of any efficiency skill this cycle โ JetBrains' benchmark walked a 30-minute session where 118K tokens of command output became 24K, validating README claims of 60โ90% savings. MD Compressor rode the "More Skills. Less Context." wave with a simple pitch: compress skill metadata so your agent can carry more capabilities per context window. context-compression-strategies rounds out the podium for long-running sessions, where context summarization is the difference between a useful agent and an amnesiac one.
| Rank | Skill | Developer | Platform | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | SkillsBench v1.1 | BenchFlow AI | skillsbench.ai | 87 native task.md packages | Major-update exemption โ and it earns it |
| ๐ฅ | Skill Eval | Minko Gechev | blog.mgechev.com + GitHub | TypeScript framework | Unit tests for skill-following |
| ๐ฅ | Agent Evaluation | sickn33 | GetClaudeSkills | SKILL.md framework | Benchmark LLM agents systematically |
Why these ranked here: SkillsBench appeared on last week's leaderboard, but the rules allow major-version exemptions โ and v1.1 shipping 87 native BenchFlow task.md packages is a genuine leap for the first framework measuring whether skills actually improve performance. Minko Gechev takes silver with Skill Eval, a TypeScript framework benchmarks how well an agent follows your skill โ define a task, point it at the skill, measure. Agent Evaluation (sickn33) rounds out a category that's becoming the ecosystem's immune system.
Ranked by AgentSkillReview. Install data from skills.sh (fetched September 11, 2026, ~09:15 CST); launch and benchmark data from vendor docs, JetBrains' rtk benchmark, BrowserAct, VoltAgent's awesome-agent-skills catalog, and community forums. Skills rotate: nothing appears in consecutive windows without a major update or a title defense.
