๐ Weekly Skill Leaderboard โ November 20, 2026
๐ Weekly Skill Leaderboard โ November 20, 2026
Governance week. NVIDIA entered the skills ecosystem in force โ four repos in one sweep โ Google shipped official skills for its own products, Jane Street open-sourced a single obsessively-crafted tutor skill that rocketed from zero to 9,732 stars in two weeks, and Cloudflare's security-audit skill delivered on last week's headliner promise with the fastest sustained velocity we've ever tracked. The catch: our 30-day rotation benched 77 previously ranked skills, so every podium slot this week went to genuinely fresh blood. Six categories, 18 skills, zero reruns.
๐ Category Rankings
๐งโ๐ป Coding & Development
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | NVIDIA/skills | NVIDIA | new (Nov entry) | SKILL.md packs | Big-hardware entrant |
| ๐ฅ | google/skills | ~22K (coddykit, Mar 2026 repo) | SKILL.md packs | The search giant's official stack | |
| ๐ฅ | Understand-Anything | Egonex-AI | 92,181 (+295/day) | SKILL.md | Watch-list graduate |
Why these ranked here:
- NVIDIA/skills is the week's biggest signal: the company that makes the silicon is now publishing Agent Skills for its own stack โ CUDA-X libraries, Physical AI and robotics workflows, simulation, and more, explicitly installable into Claude Code and Codex (github.com/NVIDIA/skills, updated within days of our research pass). When the biggest hardware vendor on the planet standardizes on SKILL.md, the format question is settled.
- google/skills landed the same week: official Agent Skills for Google products and technologies including Google Cloud (github.com/google/skills). A 22K-star Google skills framework (created March 2026, per coddykit's November trending coverage) crossing into official product repos is the second mega-vendor validation in seven days.
- Understand-Anything (Egonex-AI) graduates from our watch list: 92,181 stars at +295/day (linkly.ai, Nov 12). Codebase comprehension skills are quietly becoming the third pillar of the coding stack alongside review and planning.
๐ฃ Harness & Orchestration
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | agent-browser | vercel-labs | 49,866 (+659/day) | CLI + SKILL.md | The universal actuator |
| ๐ฅ | find-skills | vercel-labs | 4.4M installs | SKILL.md | The meta-skill |
| ๐ฅ | Skills Gateway | LiteLLM | self-hostable OSS | SKILL.md distribution | Routing infra for the format |
Why these ranked here:
- agent-browser owns browser control for agents: open pages, read them as an accessibility tree, act on elements from the shell (github.com/vercel-labs/agent-browser). 49,866 stars at +659/day (linkly, Nov 12) and the #3 skill by all-time installs on skills.sh (1.6M). If your agent touches the web, this is the current default.
- find-skills is the unlikely #1 on skills.sh's all-time leaderboard โ 4.4 million installs (skills.sh). A skill whose entire job is routing agents to other skills. Discovery infrastructure has become the single most-installed capability in the ecosystem, which tells you how fast the catalog is growing.
- LiteLLM's Skills Gateway closes the loop on the model side: upload SKILL.md packages with scripts and reference files, install them across Codex, Claude Code, and other Agent-Skills-format agents (docs.litellm.ai). Open-source and self-hostable โ gateway infrastructure adopting the format, not a SaaS walled garden.
๐จ Design & Creative
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | design-mobile-apps | designed-by-ai | 1.5M installs | SKILL.md suite | Install-mass leader |
| ๐ฅ | ui-taste | uizze.sh | 1.1M installs | SKILL.md | The taste-maker |
| ๐ฅ | web-design-guidelines | vercel-labs | in official collection | SKILL.md | Enterprise-grade baseline |
Why these ranked here:
- design-mobile-apps (designed-by-ai/skills) sits at #6 all-time on skills.sh with 1.5M installs โ the highest-installed design skill in the ecosystem right now. Last week's design podium (ui-ux-pro-max, taste-skill, hallmark) got rotation-benched; this is what the install counts look like when you zoom out to all-time demand.
- ui-taste (uizze.sh) at #18 with 1.1M installs is the anti-slop darling of the long tail: taste calibration for agents so output doesn't look AI-generated. Smaller repo, enormous per-skill pull.
- web-design-guidelines from Vercel's official agent-skills collection (SKILL.md confirmed in-repo) is the enterprise baseline โ the skill you install when the client asks "who wrote the design rules?"
๐ฌ Media Generation
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | awesome-genmedia/skills | EachLabs | new (Nov entrant) | SKILL.md + plugin | The 431-model megaphone |
| ๐ฅ | ai-image-generation | magentosh/superpowers | 956.4K installs | SKILL.md | Volume leader |
| ๐ฅ | Generative Media Skills | podo (muapi.ai) | โ | SKILL.md bundle | The ad-crew specialist |
Why these ranked here:
- awesome-genmedia/skills is the boldest swing of the week: generate images, videos, audio, and 3D models across 431 AI models through a single API, installable as a Claude Code plugin (github.com/awesome-genmedia/skills). Model-agnostic media via the SKILL.md interface is the category's logical endgame.
- ai-image-generation (magentosh/superpowers) pulls 956.4K installs on skills.sh (#28 all-time) โ proof that focused single-purpose media skills still beat big bundles for raw usage, a pattern the SkillsBench research backs (focused โค3-module skills outperform exhaustive ones; see Trends below).
- podo's Generative Media Skills take the production angle: cinematic video direction, Nano Banana-class image generation, Seedance 2.0 video, and UGC/product ad pipelines distributed via TypingMind (typingmind.com/skills/podo-generative-media-skills). Less catalog, more craft.
๐ Security & Governance
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | security-audit-skill | Cloudflare | 70,370 (+1,048/day) | SKILL.md, MIT | Top skill overall |
| ๐ฅ | SkillSpector | NVIDIA | new (active) | Scanner CLI | Supply-chain shield |
| ๐ฅ | skills | Trail of Bits | security-researcher grade | SKILL.md methodology | Professional audit doctrine |
Why these ranked here:
- Cloudflare's security-audit-skill is this week's #1 overall โ and the fastest sustained riser we've ever logged: 70,370 stars at +1,048/day (linkly.ai, Nov 12), up from ~55,800 on Nov 1 (vibe-eval), with 36,000 of its lifetime stars arriving in the last 30 days โ 57% of its entire history (community tracking). The design justifies the hype: six phases, reconnaissance through candidate verification, isolated hunter agents, and findings filed only after an independent agent tried and failed to disprove them. It headlines exactly as promised last week.
- NVIDIA/SkillSpector is the supply-chain response: a security scanner for agent skills that detects vulnerabilities, malicious patterns, prompt injection, and data exfiltration before you install (github.com/nvidia/skillspector, pushed days before this board). Snyk's ToxicSkills audit found 1,467 malicious payloads across 3,984 scanned skills โ tooling like this is no longer optional.
- Trail of Bits' skills round out the podium with what professional security researchers actually use: CodeQL and Semgrep audit methodology packaged as SKILL.md doctrine. The category returns to the board with its strongest three-way field yet.
๐ฌ Research & Evaluation
| Rank | Skill | Developer | Stars | Format | Verdict |
|---|---|---|---|---|---|
| ๐ฅ | agent_skills (socratic-tutor) | Jane Street | 9,732 (+659 forks) | single SKILL.md, MIT | New entry of the week |
| ๐ฅ | digital-health-skills | NVIDIA | new (Nov entrant) | SKILL.md + AGENTS.md | Domain-depth pioneer |
| ๐ฅ | SkillEvaluator | NVIDIA | new (open CLI + PyPI) | Evals + PyPI package | The trust instrument |
Why these ranked here:
- Jane Street's agent_skills is the most interesting 9,732-star repo in the ecosystem right now: one skill โ
socratic-tutorโ and nothing else (GitHub API Nov 17: 9,732 stargazers, 659 forks, pushed Nov 3). It encodes strict Socratic pedagogy: decompose topics into 4โ10 subskills, run a 3โ5 question diagnostic, render an ASCII mastery map, ask exactly one question per turn, refuse to bump levels without two independent checks including one unprompted recall. Anti-sycophancy as executable doctrine. Quant finance's teaching culture, shipped as a portable SKILL.md. - NVIDIA/digital-health-skills extends the format into healthcare AI workflows โ evaluation, data generation, model adaptation, and deployment guidance, with a formal AGENTS.md (github.com/NVIDIA/digital-health-skills). Skills going vertical-industry is the next adoption curve.
- NVIDIA/SkillEvaluator is how the ecosystem grows up: an open CLI and PyPI package that grades skills in tiers โ static checks, deduplication, then live agent runs with and against the skill to measure its actual effect on task results (github.com/NVIDIA/skillevaluator, explainx guide). "Does this skill actually help?" finally has a default answer.
๐ Trends This Week
- The mega-vendors have arrived. NVIDIA (4 repos), Google (official product skills), Cloudflare, Vercel, Jane Street โ every tier of the industry now ships SKILL.md. The open standard's donation to the Agentic AI Foundation (Oct 22) looks in hindsight like the starting gun.
- Skills measurably work โ now with numbers. SkillsBench's 87-task evaluation found curated skills lift average agent pass rate from 33.9% to 50.5% (+16.6pp, per-configuration gains ranging +4.1 to +25.7pp), and that focused skills (โค3 modules) beat exhaustive bundles (arXiv 2602.12670). Small, sharp, verified โ that's the winning shape.
- Trust tooling is the new frontier. Snyk's 3,984-skill audit (1,467 malicious payloads), NVIDIA's SkillSpector and SkillEvaluator, and Cloudflare's audit skill all point the same direction: the verify-before-install era has begun.
๐ Movers & Shakers
- Biggest riser: Cloudflare security-audit-skill โ +1,048 stars/day sustained; 57% of its lifetime stars arrived in the past 30 days. Nobody else is close.
- New entry of the week: Jane Street โ 0 โ 9,732 stars in 14 days with a single skill. Proof that craft beats catalog size.
- Rotation watch: the bench is terrifyingly deep โ 77 skills logged in the past 30 days are blocked, including rea (452,749โ ), ECC (301,370โ ), Caveman (119,137โ ), hyperframes (82,909โ ), hallmark, superpowers (316,301โ ), and impeccable (108,305โ ). All champions return Nov 30. Next Friday's board will be a war.
๐ฎ Next Week's Watch List
- The Nov 30 rotation release โ superpowers, rea, ECC, Caveman, hyperframes, hallmark, and impeccable all become eligible again. Expect returning champions to collide with this week's mega-vendor entrants for the first time.
- NVIDIA/skills velocity โ if the CUDA-X and robotics skill packs compound at Jane Street's pace, they'll podium next week on raw growth alone.
- reddit-automation (flowkit-labs/skills) โ #8 all-time on skills.sh at 1.5M installs; if the automation-harness category keeps climbing, it earns a board slot of its own.
Ranked by AgentSkillReview. Only standard SKILL.md-interface skills are ranked โ no platforms, no SaaS. Data sourced from GitHub, skills.sh, linkly.ai/skills/top-repos (Nov 12 UTC), explainx.ai, Snyk ToxicSkills, SkillsBench (arXiv), and community forums. Star counts are dated source snapshots, not live API readings. Rotation rule: no skill re-ranks within 30 days unless defending #1 or shipping a major update.