The top 3 modular agent skills in every category โ ranked by community signal, real-world utility, and momentum over the last 7 days. This week marks a turning point: big tech has arrived. Google, NVIDIA, and Matt Pocock all shipped major updates this month, and the ecosystem is unrecognizable from even 30 days ago.
If you're new to the modular skills ecosystem, this is your weekly cheat sheet for what's worth installing. Every skill here follows the open SKILL.md standard โ a folder with a SKILL.md (YAML frontmatter + Markdown instructions), optionally bundled with scripts/, references/, and assets/. They run across Claude Code, Codex, Cursor, Gemini CLI, Copilot, and 40+ other agents.
This week covers 7 categories ร 3 ranks = 21 skills. We hunted these across GitHub trending, skills.sh, LobeHub (333K+ skills indexed), NVIDIA's verified pipeline, Google Cloud's official repo, and community forums. Notable: all 44 skills from the previous two weeks (Aug 9 daily + Aug 14 weekly) are inside our 30-day dedup window, so this week's leaderboard is 100% fresh. One exception is flagged below where a major version update clears it.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | Matt Pocock Skills v1.0 | Matt Pocock | mattpocock/skills | 135K โญ / 11.7K forks | SKILL.md + scripts | 63% token reduction via progressive disclosure |
| ๐ฅ | Firecrawl Agent Skill | Firecrawl | firecrawl/agent-skills | Growing rapidly | SKILL.md + CLI | 80%+ content recall, 16 workflow skills |
| ๐ฅ | Skill Creator | Anthropic | anthropics/skills | Official | SKILL.md generator | Interactive Q&A โ complete skill |
๐ฅ Matt Pocock Skills v1.0 โ The first semver major from the skills ecosystem's most-watched independent developer. Version 1.0 shipped in June 2026 with a headline feature: 63% lower token costs via progressive disclosure โ compact index metadata loads at startup (~100 tokens per skill), full instruction bodies load only on skill invocation. The collection spans handoff, domain-modeling, codebase-context, routing, and behavioral guardrails. At 135K stars and 11.7K forks, it's the second most popular skills repo after addyosmani/agent-skills. We're invoking the major version update dedup override: the individual grill-me skill was reviewed Aug 9, but v1.0's architectural rewrite (progressive disclosure across every skill) is a new product. MIT-licensed, free, installable via npx skills add mattpocock/skills.
๐ฅ Firecrawl Agent Skill โ The most comprehensive web data toolkit available as an agent skill. Firecrawl teaches Claude Code (and any SKILL.md-compatible agent) to install and use its CLI autonomously: firecrawl scrape for clean markdown from any page, firecrawl search for web search with scraped results, firecrawl interact for browser automation, and firecrawl crawl for recursive site mapping. Delivers 80%+ content recall on benchmark evaluations โ outperforming every other scraping tool tested. The --all flag installs to every detected AI coding agent on your machine. Free tier (1,000 credits/month) covers testing; paid plans for production. 16 workflow skills span market research, SEO audits, competitive intelligence, and design system cloning.
๐ฅ Skill Creator (Anthropic) โ Building custom skills used to mean staring at a blank SKILL.md file and guessing at structure. Anthropic's Skill Creator fixes that with an interactive Q&A flow that surfaces edge cases you wouldn't have thought of and generates a complete, properly-structured skill directory. Available at claude.com/plugins/skill-creator or via npx skills add anthropics/skills --skill skill-creator. It asks about the workflow, trigger conditions, scripts needed, and edge case handling โ then outputs a production-ready SKILL.md with correct frontmatter and progressive disclosure structure. This is the skill that makes the skills ecosystem self-sustaining.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | Google Cloud Skills | google/skills | 18.6K โญ / 1.5K forks | SKILL.md + plugins | 90+ skills, full GCP surface | |
| ๐ฅ | Terraform & OpenTofu Skill | Anton Babenko | antonbabenko/terraform-skill | 2.3K โญ / 203 forks | SKILL.md + references | Multi-cloud testing, modules, CI/CD |
| ๐ฅ | NVIDIA Agent Skills | NVIDIA | nvidia/skills | 3K โญ / 352 forks | SKILL.md + verified pipeline | 150+ skills, Physical AI to CUDA |
๐ฅ google/skills โ Launched at Cloud Next 2026 with 13 skills; now at 90+ and growing daily. This is the most comprehensive cloud-native skills repository from any major vendor. Covers the full GCP surface: AI/ML (Agent Platform, Gemini API, BigQuery AI), infrastructure (GKE with 20+ dedicated skills, Cloud Run, networking), databases (AlloyDB, Spanner, Bigtable, Cloud SQL), security (IAM, PAM, SecOps), and even advertising (Google Ads API, Mobile Ads SDK). Each skill is maintained by the product team that owns the underlying service. Apache 2.0 license. Install via npx skills add google/skills. The repo also bundles plugins for Claude Code, Codex, and Antigravity CLI โ skills + MCP servers in one install.
๐ฅ antonbabenko/terraform-skill โ The definitive IaC skill for AI agents. Created by the maintainer of terraform-aws-modules and terraform-best-practices.com, this skill gives agents field-tested patterns for testing (native tests vs Terratest decision matrix), module development (naming, versioning, public/private patterns), state management (S3/Azure/GCS backends with locking and encryption), CI/CD (GitHub Actions, GitLab CI, Atlantis, Infracost), and security (Trivy, Checkov, policy-as-code). AWS, Azure, and GCP are all first-class โ ask for the Azure or GCP equivalent of any pattern and the skill maps it. 2.3K stars reflect the Terraform community's trust in Babenko's work. Apache 2.0.
๐ฅ nvidia/skills โ The most ambitious skills catalog from any hardware vendor. 150+ skills span the full NVIDIA stack: Physical AI (Omniverse CAD-to-sim-ready, neural reconstruction, defect image generation), robotics and simulation (Isaac Sim, Holoscan, HoloHub), CUDA-X libraries (cuDF, cuOpt, cuPyNumeric), RAG and AI Blueprints, healthcare (DICOM, MONAI, clinical ASR), Earth2Studio for weather/climate AI, and Jetson for edge deployment. Every skill goes through NVIDIA's verified pipeline โ scanned, evaluated, and signed by SkillSpector before publication. The catalog is maintained via an automated daily sync from product repos. 582 commits, 3K stars in just weeks since launch. Free, open-source (Apache 2.0 / CC-BY-4.0).
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | Remotion Best Practices | Remotion Team | remotion-dev/skills | 117K weekly installs | SKILL.md + rules | Programmatic video with React, full stack |
| ๐ฅ | Marketing Skills (32-skill suite) | Corey Haines | coreyhaines31/marketingskills | 12.9K โญ / 1.9K forks | SKILL.md + context files | Full-funnel marketing automation |
| ๐ฅ | NVIDIA Physical AI Skills | NVIDIA | nvidia/skills | 3K โญ | SKILL.md + verified | Synthetic data, simulation, visual AI |
๐ฅ Remotion Best Practices โ With 117K weekly installs across Claude Code, Codex, Gemini CLI, and Cursor, this is one of the most widely adopted official skills in the ecosystem. Maintained by the Remotion team, it loads specialized rules for animations (interpolation curves, spring animations, easing, sequencing, transitions), audio and captions (importing, trimming, volume control, subtitles), media handling (video, images, GIFs, Lottie, fonts, transparent rendering), 3D content (Three.js + React Three Fiber inside Remotion), charts and data viz (bar, pie, line, stock charts), and advanced patterns (dynamic metadata, parametrizable videos with Zod schemas, ElevenLabs voiceover). Activates automatically when Claude detects Remotion code; only loads relevant rule files on demand to stay token-efficient.
๐ฅ Corey Haines Marketing Skills โ 32 skills organized into a complete marketing stack: conversion optimization (page-cro, signup-flow-cro, onboarding-cro), copywriting (homepage, landing, feature copy), SEO (seo-audit, programmatic-seo, ai-seo for LLM answer optimization), paid ads, analytics (analytics-tracking, GA4), retention (churn-prevention, email-sequence, cold-email), and sales operations (revops). All skills pull from a shared product-marketing-context file โ fill in your product details once and every skill gets context-aware. 12.9K stars reflect the hunger for marketing skills in an agent-driven world. Free, install selectively: npx skills add coreyhaines31/marketingskills --skill page-cro copywriting seo-audit.
๐ฅ NVIDIA Physical AI Skills โ A subset of NVIDIA's catalog that deserves its own spotlight in the creative category. These skills teach agents to generate synthetic training data (physical-ai-defect-image-generation, physical-ai-video-data-augmentation, physical-ai-image-attribute-augmentation), run neural reconstruction from real-world captures (physical-ai-neural-reconstruction), and set up Omniverse pipelines (omniverse-cad-to-simready, omniverse-realtime-viewer, omniverse-usd-performance-tuning). They bridge the creative-media and industrial-simulation worlds โ ideal for teams building AI-powered visual pipelines. NVIDIA-verified, free.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | NVIDIA AIQ Research | NVIDIA | nvidia/skills | 3K โญ | SKILL.md + deploy | Shallow + deep research workflows |
| ๐ฅ | Google Agent Platform Eval Flywheel | google/skills | 18.6K โญ | SKILL.md | Continuous eval improvement loop | |
| ๐ฅ | NeMo Retriever | NVIDIA | nvidia/skills | 3K โญ | SKILL.md + scripts | RAG extraction, deploy, Q&A |
๐ฅ NVIDIA AIQ Research โ NVIDIA AI-Q Blueprint exposed as two agent skills: aiq-research for running shallow and deep research workflows, and aiq-deploy for deploying local AI-Q services. The deep research workflow forces multi-angle exploration, dimension-specific deep dives, and diversity/validation checks โ a methodology reminiscent of ByteDance's deer-flow but packaged for the NVIDIA ecosystem. This is the first major hardware vendor to ship research methodology as a modular agent skill. Free, install via npx skills add nvidia/skills --skill aiq-research.
๐ฅ Google Agent Platform Eval Flywheel โ Continuous evaluation as an agent skill. This Google Cloud skill implements an eval flywheel: define evaluation criteria โ run agent tasks โ score outputs โ identify gaps โ improve prompts/skills โ repeat. It's the most structured approach to "how do I know if my agent is getting better or worse over time" we've seen packaged as a single SKILL.md. Part of Google's Agent Platform suite within the google/skills repository. Apache 2.0.
๐ฅ NeMo Retriever โ NVIDIA's RAG skill: deploy the NeMo Retriever Library locally, extract information from a corpus of data, and answer questions against that corpus. It's a Capability Uplift skill that turns any SKILL.md-compatible agent into a RAG operator โ handling embedding, retrieval, and synthesis. Free.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | Handoff | Matt Pocock | mattpocock/skills | 135K โญ | SKILL.md | Structured context transfer across agents |
| ๐ฅ | NeMo Relay | NVIDIA | nvidia/skills | 3K โญ | SKILL.md + plugins | Runtime for agent instrumentation & control |
| ๐ฅ | Google Agent Platform Endpoint Management | google/skills | 18.6K โญ | SKILL.md | Managed agent endpoints at scale |
๐ฅ Handoff (Matt Pocock) โ The structured answer to "my session is hitting 120K tokens and getting dumber." /handoff generates a markdown document containing the purpose of the next session, relevant context, suggested skills to invoke, and pointers to existing artifacts โ without duplicating file content. The key insight: it's designed for cross-agent transfer, not just session continuation. Plan in Claude Code, hand off to multiple Codex worktrees for parallel implementation. As Ben Holmes demonstrated, this pattern turns a single planning session into N parallel implementation sessions. Part of Matt Pocock Skills v1.0; free, MIT.
๐ฅ NeMo Relay (NVIDIA) โ A runtime for instrumenting and controlling AI agents across harnesses, applications, and frameworks. Skills cover installation (nemo-relay-install), getting started (nemo-relay-get-started), instrumenting calls, context isolation, typed wrappers, plugin development (adaptive tuning, observability, custom builds), migration from flow-based systems, and debugging runtime integration. Think of it as an observability and control plane for multi-agent systems, packaged as modular skills. Free.
๐ฅ Google Agent Platform Endpoint Management โ Google's skill for managing agent endpoints at production scale: deployment, versioning, traffic splitting, monitoring, and rollback. Part of the google/skills Agent Platform suite, it's the operational backbone skill that turns a prototype agent into a managed service. Apache 2.0.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | SkillSpector | NVIDIA | nvidia/skillspector | New (Aug 20) | CLI + verification | Scans, evaluates, signs agent skills |
| ๐ฅ | Trail of Bits Security Skills | Trail of Bits | trailofbits/skills | Professional | SKILL.md + CodeQL/Semgrep | Professional-grade static analysis |
| ๐ฅ | NemoClaw | NVIDIA | nvidia/skills | 3K โญ | SKILL.md + OpenShell | Secure agent sandboxing |
๐ฅ SkillSpector (NVIDIA) โ Launched just yesterday (Aug 20, 2026), SkillSpector is part of NVIDIA's Verified Skills pipeline. It scans, evaluates, and cryptographically signs agent skills before publication. Skills that pass receive an NVIDIA-verified signature, giving enterprises a trust signal in an ecosystem where Snyk's ToxicSkills research found prompt injection in 36% of tested skills. SkillSpector can also be run in GitHub Actions for CI/CD integration โ scan your team's skills on every PR. This is the governance infrastructure the ecosystem has been waiting for. Updated 1 day ago, this is the freshest security skill on the board. Free, open-source.
๐ฅ Trail of Bits Security Skills โ From the security research firm behind Slither, Echidna, and countless high-profile audits. These skills bring professional-grade static analysis into any SKILL.md-compatible agent: CodeQL and Semgrep analysis, variant analysis for finding related vulnerabilities across a codebase, and structured code auditing following professional audit methodology. These aren't basic security checklists โ they encode the actual workflows Trail of Bits uses for paid engagements. Requires CodeQL or Semgrep installed locally. Free.
๐ฅ NemoClaw (NVIDIA) โ Secure agent sandboxing that runs OpenClaw inside NVIDIA OpenShell with managed inference, policy management, remote deployment, and sandbox monitoring. The skill (nemoclaw-user-guide) teaches agents to operate within governed sandboxes โ a critical capability as enterprises move from "can my agent do X" to "can my agent do X safely." Free.
| Rank | Skill | Developer | Platform | Stars | Format | Verdict |
|---|---|---|---|---|---|---|
| ๐ฅ | SkillReducer | Research (arXiv) | arXiv 2603.29919 | Academic | Methodology | Empirical token optimization for skills |
| ๐ฅ | Matt Pocock Progressive Disclosure | Matt Pocock | mattpocock/skills | 135K โญ | Architecture pattern | 63% token reduction, compact index |
| ๐ฅ | Context Mode | Julius Brussee | Community | Active | SKILL.md | Shell output filtering + state restoration |
๐ฅ SkillReducer โ Published on arXiv (2603.29919), this is the first academic paper to tackle the problem of skill token bloat head-on. SkillReducer removes non-essential content from skills to reduce token cost, with each empirical finding driving a specific reduction strategy. While it's a research methodology rather than a drop-in SKILL.md, its impact on the ecosystem is already visible: the paper's findings (that 40-60% of typical SKILL.md content is non-essential for the agent to perform the task) directly inform the progressive disclosure patterns now being adopted by Matt Pocock, Anthropic, and others. Free/open access.
๐ฅ Matt Pocock Progressive Disclosure โ The v1.0 architecture that delivers 63% token reduction deserves its own spot on this leaderboard, separate from the Coding category. The pattern: compact index metadata (~100 tokens per skill at startup), full instruction bodies only on invocation, and reference files loaded on demand. It's not just a feature โ it's becoming the reference architecture for how skills should be structured to minimize context pollution. This is the implementation that proves SkillReducer's academic findings work in production. MIT, free.
๐ฅ Context Mode โ Companion to the Caveman skill (which is in our 30-day dedup). Context Mode filters shell output noise before it reaches the context window โ every git status, npm test, and docker build command that would otherwise dump raw output into context gets trimmed to what matters. It also restores full session state when Claude resets, meaning you don't lose your place after a compaction event. Together with progressive disclosure, this represents the "defense in depth" approach to token efficiency. Free, open-source.
Ranked by AgentSkillReview. Data sourced from GitHub, skills.sh leaderboard (133K+ skills), LobeHub (333K+ indexed), NVIDIA verified pipeline, Google Cloud official repo, arXiv, and community forums. Install counts and stars verified against live sources Aug 17โ21, 2026.
Methodology: (1) community signal โ GitHub stars, skills.sh installs, weekly growth; (2) recency โ commits in past 30 days; (3) modularity โ clean SKILL.md + optional scripts/references; (4) real-world utility; (5) innovation. All skills follow the open SKILL.md standard. 30-day dedup enforced against reviewed_skills. No internal skills reviewed. One dedup override granted for Matt Pocock Skills v1.0 (major version with architectural rewrite).
