TL;DR — The short version
Google's entry into agent skills is a genuinely massive story. In March 2026, Google open-sourced google/skills (16,350★, Apache-2.0) — a first-party skills repo for Google Cloud, AI/ML, and Infrastructure — and it passed 16K stars in roughly four months. The market now splits into three format camps: Anthropic's .agents convention (mattpocock, superpowers), Google's .skills distribution model, and the cross-harness community .md movement (SKILL.md, backed by agentskills.io and the skills.sh installer). For most developers, obra/superpowers (268,868★) remains the strongest all-round methodology, while google/skills is the fastest-rising tool for anyone already in the Google Cloud / Gemini ecosystem.
01 Why this matters: Google just legitimized agent skills
For most of 2025, "skills" were a community phenomenon: markdown folders that teach AI coding agents how to do specific tasks. Anthropic formalized them as Agent Skills (a SKILL.md folder with metadata, instructions, and scripts), and the ecosystem exploded — obra/superpowers (268,868★), mattpocock/skills (209,082★), and addyosmani/agent-skills (84,035★) became some of the fastest-growing repos in GitHub history.
Then, in March 2026, Google entered the arena with google/skills — an official, Apache-2.0 repository of Agent Skills for Google products and technologies. That single move changed the competitive landscape in three ways:
- It legitimized the category. When a hyperscaler ships first-party skills for its cloud, the format stops being a hack and becomes a platform surface.
- It made the format war three-way. Anthropic owns the .agents convention and the SKILL.md standard; Google is pushing .skills distribution tied to Gemini CLI, Vertex AI, and Google Cloud; and the community keeps shipping plain .md skills that run everywhere.
- It created a cloud-lock-in surface. Skills that teach agents "how to do X on Google Cloud" are quietly the new vendor-specific documentation — for Gemini, Azure (microsoft/skills), and Anthropic (anthropics/skills) alike.
Because this category is hours old in SEO terms, there is essentially zero existing comparison content for "Google agent skills" — this page is the first comprehensive ranking of the full ecosystem.
02 The 10 best agent skills frameworks, ranked by GitHub stars
All star counts are live GitHub data collected August 8, 2026. The ecosystem splits into four camps: Google Anthropic Microsoft Community
| # | Tool / Repo | Stars | What it is | Best for | Deployment |
|---|---|---|---|---|---|
| 1 | obra/superpowers | ★ 268,868 | Agentic skills framework & full software-development methodology | End-to-end SDLC with subagent-driven development | Local install into agent harness |
| 2 | mattpocock/skills | ★ 209,082 | Real-engineer skills straight from a production .agents directory |
Battle-tested engineering workflows | .agents/ directory pattern |
| 3 | anthropics/skills | ★ 166,923 | Official Agent Skills repo — reference SKILL.md implementations | Learning the standard + canonical examples | SKILL.md folders (skills.sh) |
| 4 | addyosmani/agent-skills | ★ 84,035 | Production-grade engineering skills for AI coding agents | Cross-harness engineering workflows | Portable skills, 5+ harnesses |
| 5 | wshobson/agents | ★ 38,605 | Multi-harness agentic plugin marketplace | Discovering & installing skills across agents | Marketplace + installer |
| 6 | zhaoxuya520/reverse-skill | ★ 20,663 | AI-powered cybersecurity skill router with self-evolving knowledge base | Security research, RE, pentesting | Router pack + toolchain bootstrap |
| 7 | google/skills | ★ 16,350 | Official Google Agent Skills for Google Cloud, AI/ML, Infrastructure | Google Cloud / Gemini / Vertex work | npx skills add google/skills |
| 8 | Doriandarko/maestro | ★ 4,355 | Framework for intelligently orchestrating subagents with skills | Subagent orchestration & delegation | Python framework (pip) |
| 9 | microsoft/skills | ★ 2,868 | Skills, MCP servers, custom agents & Agents.md for SDKs | Azure SDKs / Microsoft AI Foundry | SDK packages (skills.sh) |
| 10 | microsoft/skill-recorder | ★ 2,378 | Desktop app that records your screen and reconstructs a reusable Skill | Creating skills by recording real work | Electron desktop app |
Star counts fetched live from the GitHub REST API on August 8, 2026. Ranks change daily — this ecosystem is growing fast.
03 Deep dive: every tool in the ecosystem
1. obra/superpowers
★ 268,868What it is: "An agentic skills framework & software development methodology that works." The single most-starred repo in the ecosystem — a complete skills library (brainstorming, planning, coding, debugging) fused with a subagent-driven development methodology (SDLC).
- Full SDLC methodology, not just individual skills
- Subagent-driven development with delegation workflows
- Battle-tested brainstorming → planning → coding loop
- Huge, active community & rapid iteration
2. mattpocock/skills
★ 209,082What it is: "Skills for Real Engineers. Straight from my .agents directory." Matt Pocock published his personal, production-grade skills — the definitive demonstration of the .agents directory pattern that Anthropic-style workflows popularized.
- Real skills used in production by a prominent engineer
- Reference implementation of the
.agents/layout - Focus on practical, reusable engineering tasks
- Simple, copy-paste friendly structure
3. anthropics/skills
★ 166,923What it is: The official Agent Skills repository from Anthropic — the reference implementation of the SKILL.md standard (agentskills.io). Includes creative, technical, and enterprise skills, plus the document skills (docx, pdf, pptx) that power Claude's document capabilities.
- Canonical SKILL.md format reference
- Document creation & editing skills (docx/pdf/pptx)
- Cross-domain examples: art, music, design, testing, MCP
- Powers Claude.ai's built-in document features
4. addyosmani/agent-skills
★ 84,035What it is: "Production-grade engineering skills for AI coding agents" from Addy Osmani (Google Chrome team). Explicitly tested across Claude Code, Cursor, Codex, GitHub Copilot, and Gemini CLI — the gold standard for portable, harness-agnostic skills.
- Engineered for 5+ agent harnesses out of the box
- Production-grade engineering workflows
- Practical, high-signal skill design
- Great documentation and install instructions
5. wshobson/agents
★ 38,605What it is: A multi-harness agentic plugin marketplace for Claude Code, Codex CLI, Cursor, OpenCode, GitHub Copilot, and Gemini CLI. The distribution layer of the ecosystem — where you discover, install, and manage skills as plugins.
- One marketplace across 6+ agent harnesses
- Curated plugin/skill registry with easy installs
- Centralized management of agent configuration
- Actively maintained by a prominent OSS developer
6. zhaoxuya520/reverse-skill
★ 20,663What it is: A cybersecurity "Skills Router" pack: AI-powered routing + on-demand toolchain bootstrapping + self-evolving knowledge base for reverse engineering, authorized penetration testing, and security research. Supports Claude Code, Kiro, Cursor, and Cline.
- AI routes tasks to the right skill automatically
- Bootstraps security toolchains on demand
- Self-evolving knowledge base (TraceCards, capability graph)
- Niche vertical depth — the anti-generic skills repo
7. google/skills
★ 16,350What it is: Google's official Agent Skills repository — the story of 2026. First-party skills for Google Cloud, AI/ML, and Infrastructure: cloud auth & onboarding recipes, full solution-architecture workflows, GKE AI migration, Agent Platform (Model Garden, endpoint management, eval flywheels), RAG on GKE/AlloyDB, and more. Installs via npx skills add google/skills.
- First hyperscaler-first-party skills repo (Apache-2.0)
- Deep Google Cloud: GKE, Vertex/Agent Platform, AlloyDB
- Solution-architecture & multi-product workflows
- Installable via skills.sh / npx, actively developed
8. Doriandarko/maestro
★ 4,355What it is: A framework for Claude Opus to intelligently orchestrate subagents — the layer above skills. Maestro manages task decomposition and delegation, and pairs naturally with skills libraries that define what each subagent should do.
- Intelligent subagent orchestration & delegation
- Task decomposition pipelines
- Complements skills libraries (skills = how, maestro = who)
- Well-documented Python framework
9. microsoft/skills
★ 2,868What it is: Microsoft's official skills repo — "Skills, MCP servers, Custom Agents, Agents.md for SDKs to ground Coding Agents." The Azure counterpart to google/skills, targeting Azure SDKs and Microsoft AI Foundry, with evals & test harnesses in CI.
- Azure SDK & AI Foundry workflows
- Skills + MCP servers + custom agents + Agents.md
- CI evals and test harness for quality
- Installable via skills.sh
10. microsoft/skill-recorder
★ 2,378What it is: A desktop app that records your on-screen work session and uses the GitHub Copilot CLI to reconstruct it as intent + ordered steps, then builds a reusable Skill or Automation for Microsoft Scout, Copilot Cowork, or Copilot Studio. Skills authoring without writing markdown.
- Record → reconstruct → publish skill pipeline
- Copilot CLI converts sessions into structured skills
- Targets Scout, Copilot Cowork, Copilot Studio
- Signals the future: skills created by demonstration
04 Key insights from the data
Google's entry legitimizes the category
google/skills hit ~16,350★ in under five months with zero marketing push — and it's Apache-2.0, installable via skills.sh, and actively developed. When a hyperscaler ships first-party skills, the format stops being a community quirk and becomes a platform surface every vendor must support.
The format war is now three-way
Anthropic pushed .agents + SKILL.md; Google is shipping .skills distribution tied to Gemini CLI and Cloud; the community keeps winning with plain .md portability. Three of the top-4 repos (superpowers, mattpocock, addyosmani) are format-agnostic markdown — portability is the real battleground.
Stars are massively concentrated at the top
superpowers (268,868) + mattpocock (209,082) + anthropics (166,923) + addyosmani (84,035) hold ~80% of the ecosystem's stars. The long tail is where niche depth lives — reverse-skill (20,663★) proves a focused vertical can still explode.
Skills are becoming a cloud-lock-in surface
google/skills teaches agents "how to do X on Google Cloud"; microsoft/skills does the same for Azure/Foundry; anthropics/skills encodes Claude's capabilities. Skills are quietly replacing vendor docs as the interface between agents and clouds.
Cross-harness compatibility is the differentiator
addyosmani/agent-skills and wshobson/agents explicitly target 5–6 harnesses (Claude Code, Cursor, Codex, Copilot, Gemini CLI, OpenCode). Single-harness skills are becoming the exception, not the rule — portability now drives adoption.
Recording is the new authoring flow
microsoft/skill-recorder (Jul 2026) turns screen recordings into reusable skills via Copilot CLI. Hand-writing SKILL.md is being replaced by demonstration — the same "show, don't tell" shift that hit test automation a decade ago.
skills.sh is emerging as the package manager
google/skills, anthropics/skills, and microsoft/skills all publish skills.sh install badges. A neutral installer + registry is what makes the three-way format war survivable — the ecosystem is building its own npm.
Zero-competition SEO window is wide open
Search volume for "Google agent skills" spiked after the March 2026 launch, and virtually no comparison content exists yet. Early, data-backed pages on this category are capturing demand with essentially no ranking competition.
05 Which should you choose?
Recommendations by use case
Choose obra/superpowers. It's not just skills — it's a working SDLC (brainstorm → plan → subagent-driven build → verify) with the largest community in the space. Best starting point for teams going all-in on agent-driven development.
Choose addyosmani/agent-skills for guaranteed cross-harness portability, or mattpocock/skills for the raw .agents directory pattern from a real production workflow. Both are MIT and drop-in.
Choose google/skills. First-party Google skills for cloud auth, onboarding, GKE AI migration, Agent Platform, and RAG on AlloyDB — installed with npx skills add google/skills. This is the fastest-rising repo in the category for a reason.
Choose microsoft/skills for Azure SDK workflows (with CI evals baked in), and pair it with microsoft/skill-recorder if you want to generate skills from real recorded sessions instead of writing markdown.
Choose zhaoxuya520/reverse-skill. AI-powered routing, on-demand toolchain bootstrapping, and a self-evolving knowledge base make it the deepest vertical skills pack in the ecosystem.
Choose anthropics/skills. It's the canonical reference — including the docx/pdf/pptx skills that power Claude's document features — plus examples spanning creative, technical, and enterprise use cases.
Choose wshobson/agents as your marketplace/installer layer across Claude Code, Codex, Cursor, OpenCode, Copilot, and Gemini CLI — then layer any skills library on top.
Choose Doriandarko/maestro to handle decomposition and delegation, and let a skills library (superpowers or addyosmani) define what each subagent executes.
06 Gear up: hosting, IDE & cloud partners
Every framework above runs inside an AI coding agent. Here are the hosting, IDE, and cloud services we recommend pairing with your skills setup.
07 Frequently asked questions
What is Google Skills (google/skills)?
google/skills is Google's official open-source repository of Agent Skills for Google products and technologies — Google Cloud, AI/ML, and Infrastructure. It launched in March 2026, is Apache-2.0 licensed, and installs via npx skills add google/skills. It includes cloud auth/onboarding recipes, solution-architecture workflows, GKE AI migration guides, and Agent Platform skills.
What's the difference between .agents and .skills?
.agents is the directory convention popularized by the community and Anthropic-style workflows (a .agents/ folder in your repo containing skill folders). .skills refers to the Agent Skills distribution model — SKILL.md folders that agents load dynamically, standardized at agentskills.io and installable via skills.sh. Community .md skills are plain portable markdown that run in any harness. In practice the three overlap heavily; the battle is over which convention becomes the default.
Which agent skills framework is best for me?
For a full software-development methodology, choose obra/superpowers. For battle-tested engineering workflows, addyosmani/agent-skills or mattpocock/skills. For Google Cloud work, google/skills. For Azure/AI Foundry, microsoft/skills. For security research, reverse-skill. For learning the standard, anthropics/skills.
What is skills.sh?
skills.sh is the emerging cross-repository installer for Agent Skills. Official repos including google/skills, anthropics/skills, and microsoft/skills all publish skills.sh install badges, making it the de facto package manager for the agent skills ecosystem.
Are agent skills portable across AI coding agents?
Mostly yes. The SKILL.md markdown format is portable across Claude Code, Cursor, Codex CLI, GitHub Copilot, Gemini CLI, OpenCode, Cline, and more. Multi-harness repos like addyosmani/agent-skills and wshobson/agents explicitly test and target five or more harnesses.