Two Solo GitHub Repos Outgrow Anthropic's Own Skills Standard

Google's own entrant into the format Anthropic gave away for free is still warning contributors of "likely major breaking changes" before it reaches a stable release.

Anthropic's own reference implementation of Agent Skills carries 178,318 stars on GitHub. Two personal repositories built on the same file format have more: 291,665 for Jesse Vincent's Superpowers and 269,728 for Matt Pocock's Skills, according to GitHub's API as of Friday.

Anthropic published Agent Skills as a free, cross-platform standard on Dec. 18, 2025, at agentskills.io, and Microsoft, OpenAI, Atlassian, Figma, Cursor and GitHub adopted it the same day, with partner skills from Canva, Stripe, Notion and Zapier launching alongside it, according to Anthropic and a Unite.AI analysis of the release. Nine months later, agentskills.io's own client directory lists 46 products that support the format, from Cursor and VS Code to Snowflake's Cortex Code, according to the site as of Friday.

What Anthropic gave away when it opened Agent Skills

A skill is a folder containing a SKILL.md file: metadata plus instructions an agent loads only when a task calls for them, a design Anthropic calls progressive disclosure. Anthropic first described the format in an Oct. 16, 2025 engineering post, then handed the specification to a separate GitHub organization, agentskills/agentskills, when it opened the standard in December. That specification repository carries roughly 25,700 stars of its own, licensed Apache 2.0 for code and CC-BY-4.0 for documentation.

Anthropic's own reference implementation, anthropics/skills, is a different repository: a library of skills built for Claude, created Sept. 22, 2025. It has 21,110 forks and 1,294 open issues, and its README licenses most of its contents under Apache 2.0, with the document-editing skills marked source-available rather than open source.

The two repos with more stars than Anthropic's own

Vincent created Superpowers on Oct. 9, 2025, a week before Anthropic's first public description of Agent Skills. His GitHub profile lists his employer as Prime Radiant. The MIT-licensed repository, built mostly in Shell, describes itself as "An agentic skills framework & software development methodology that works" and has 26,108 forks.

Pocock created his Skills repository on Feb. 3, 2026, two months after the standard went cross-platform. Pocock, based in Oxfordshire, built the Total TypeScript course after leaving Vercel and Stately, according to his GitHub profile. His repository, also MIT-licensed, is subtitled "Skills for Real Engineers. Straight from my .agents directory" and links to aihero.dev/skills; it has 22,728 forks.

Between them, Vincent's and Pocock's repositories carry 561,393 stars, more than three times Anthropic's own 178,318.

Google's bet: orchestration, not the file

Google took a different position on Skills this spring. Its google/ax repository, created March 30, 2026, describes itself as "Google's open agentic orchestration runtime." It gained 1,379 stars in the 24 hours before Friday, the fastest one-day growth among the four repositories, reaching 11,504 total.

ax does not compete with SKILL.md files; it schedules them. Written in Go and licensed Apache 2.0, it declares agentic tasks with workspaces and model specifications, then sandboxes and runs them across a cluster using kubectl-style commands: apply, get, watch, delete, plus ax ssh for an interactive shell into a running task. Its own README says the project is "still actively refining core concepts, protocols, and specifications" and warns of "likely major breaking changes" before a stable release.

Why the star counts overstate the case

None of this proves adoption. A GitHub star costs a viewer one click and no commitment, and a repository can accumulate them through link-sharing without a single production deployment behind it. GitHub's trending page has rewarded attention over deployment before: Paperclip passed 80,000 stars five months after a maximum-severity vulnerability in the project became public, and Vectorize's Hindsight memory library, covered here, topped GitHub's trending page on Sept. 24 on the strength of a benchmark score first published nine months earlier.

The harder number to fake is agentskills.io's client directory, which requires a working integration and a documentation page rather than a click: 46 listed today, including GitHub Copilot, Claude Code, ChatGPT's Codex and Google's own Gemini CLI. That list is also the clearest evidence the format has moved past Anthropic: Microsoft's Copilot, OpenAI's Codex and Google's own Gemini CLI all sit in the same directory as tools nobody at any of those three companies built.

Anthropic has taken this approach before. It donated the Model Context Protocol, its standard for connecting agents to tools, to the Linux Foundation with Google, Microsoft and AWS as members, a move Unite.AI's analysis framed as competing on execution rather than on lock-in. Agent Skills follows the same shape: give away the format, keep building the product that uses it best.

1. Superpowers (Vincent)292K stars
2. Skills (Pocock)270K stars
3. anthropics/skills178K stars
4. google/ax12K stars

Source: GitHub API, accessed 2026-09-26

Whether that bet holds depends less on any single repository's star count than on what Google ships next. ax's own maintainers have not set a date for the stable release its README says is still pending, and the project has not said whether the tasks it schedules will read the same SKILL.md files that Vincent's, Pocock's and Anthropic's own repositories already define.

Sources

1. Equipping agents for the real world with Agent Skills (Anthropic)
2. Anthropic Opens Agent Skills Standard, Continuing Its Pattern of Building Industry Infrastructure (Unite.AI)
3. Agent Skills client directory (agentskills.io)
4. google/ax (GitHub)
5. obra/superpowers (GitHub)
6. mattpocock/skills (GitHub)
7. anthropics/skills (GitHub)