Agent Skills: Why Matt Pocock's 27-Skill Repo Has More Stars Than Anthropic's

A skill is a folder with a SKILL.md file in it, and two repositories full of them are climbing GitHub's trending list today: mattpocock/skills and anthropics/skills. The first shows 320.7k stars on its page and the second 192.1k, which means the individual developer's repo is ahead of the company that wrote the format.

Stars are attention, not installs. The more useful comparison is what each repo asks you to trust.

What a skill is

The Agent Skills specification defines a skill as a directory with a required SKILL.md, plus optional scripts/, references/ and assets/ folders. The file opens with YAML frontmatter. name is required, at most 64 characters, lowercase letters, numbers and hyphens, and it must match the directory name. description is required and capped at 1,024 characters.

The spec's design is progressive disclosure. An agent loads roughly 100 tokens of metadata per skill at startup, reads the full SKILL.md body, recommended under 5,000 tokens, only when it activates the skill, and pulls files from the other folders on demand. It recommends keeping SKILL.md under 500 lines.

The practical consequence is that the description line does the routing. A vague one means the agent never picks the skill; the spec's own poor example is "Helps with PDFs."

Two repos, two intents

Anthropic's repository holds example skills in four groups, the spec, and a template. The document skills for docx, pdf, pptx and xlsx are Apache 2.0, and the page says many other skills are open source under the same license, so each folder needs checking. Its README disclaimer is direct: the skills are "provided for demonstration and educational purposes only," and users should test them thoroughly before relying on them for critical tasks.

Matt Pocock's repository is MIT-licensed and describes itself as "Skills for Real Engineers. Straight from my .agents directory." It lists 27 skills, split by who triggers them.

Group	Skills	Examples
Engineering, user-invoked	11	triage, to-spec, implement
Engineering, model-invoked	9	tdd, diagnosing-bugs, code-review
Productivity, user-invoked	5	grill-me, handoff, teach
Productivity, model-invoked	2	grilling, writing-for-agents

The split matters. User-invoked skills are commands a person types to orchestrate a task. Model-invoked skills are standing discipline the agent can pick up on its own.

1. mattpocock/skills320.7 thousand stars
2. anthropics/skills192.1 thousand stars

Source: github.com/mattpocock/skills and github.com/anthropics/skills, accessed 2026-11-10

Where the friction is

Installation is the sticking point in Pocock's README. For Claude Code the command is claude plugin install mattpocock-skills@claude-plugins-official, and Codex, GitHub Copilot, Gemini CLI and others each get their own route. The page warns to use one method per agent, because installing through two duplicates every skill.

After installing, a one-time /setup-matt-pocock-skills run is needed per repository.

Anthropic's route is a marketplace: /plugin marketplace add anthropics/skills, then /plugin install document-skills@anthropic-agent-skills.

Pocock is candid about limits. He says /improve-codebase-architecture is "a survey, not a rescue" and will not untangle a messy codebase alone. His page also argues that frameworks such as GSD, BMAD and Spec-Kit "take away your control." That is his opinion, and the repo carries 210 open issues and 22 pull requests, with PRs limited to collaborators.

What to check before installing

A skill is instructions plus scripts that an agent may run with your permissions. Read the SKILL.md and anything in scripts/ first, and check the allowed-tools field, which the spec marks experimental and says may vary between agent implementations.

For related tooling coverage, see agent-browser 0.42 and Recalc, the spreadsheet for Claude Code and Codex.

The thing to watch next is whether the spec's compatibility and allowed-tools fields get consistent support across agents, since a skill that works in one tool and is ignored in another is the likeliest source of confusion.