S11 Six months of Anthropic Skills
Claude Skills,
mapped.
Anthropic announced Agent Skills on October 16, 2025. Six months later, the ecosystem has seventeen official skills, twenty-two vendor packs, tens of thousands of community skills, and a cross-vendor open standard called SKILL.md. The whole arc fits in two quarters and four charts.
A skill is a folder with a YAML-fronted Markdown file and optional scripts. The format is so simple that within sixty days of announcement, OpenAI, Cursor, Gemini, Antigravity, Windsurf, and OpenCode all adopted it. Below: the surge timeline, the power-law star distribution, the architectural sweet spot, and the verbs every skill author opens with.
§ I S11.1 · The arc, in one paragraph
Two quarters. One open standard.
In September 2025, Claude.ai quietly shipped the ability to create Word docs, PDFs, slide decks, and spreadsheets. Few people noticed it was powered by a new internal abstraction. In October, Anthropic announced the abstraction publicly. By December, they had open-sourced the format. By March 2026, Cursor, OpenAI's Codex CLI, Gemini CLI, and three others were shipping skills in the same SKILL.md syntax.
A Claude Skill is a folder. Inside it is a SKILL.md file with
YAML frontmatter (a name, a description,
optional allowed-tools) and a Markdown body. Optional
subfolders are scripts/, references/,
and assets/. The headline design choice is
progressive disclosure: at session start, only the YAML
frontmatter loads, costing about a hundred tokens per skill.
The Markdown body loads only when a task matches. Bundled files
load only when explicitly read. Anthropic's API caps requests at
eight skills and eight megabytes per skill.
That's the format. The phenomenon is the surge. Three audiences adopted it simultaneously: developers who want workflow orchestration in Claude Code, knowledge workers who want document creation in Claude.ai, and vendors who want their product to be the default tool an agent reaches for. Tens of thousands of skills shipped in the first six months. The four charts below tell the rest of the story.
- Sep 2025
- Document creation (docx/pdf/pptx/xlsx) ships in Claude.ai, secretly powered by skills
- Oct 16, 2025
- Anthropic publicly announces "Agent Skills" via engineering blog and anthropics/skills repo
- Oct 9, 2025
- Jesse Vincent (obra) publishes the Superpowers framework
- Oct 16-18, 2025
- Simon Willison's "maybe a bigger deal than MCP" post hits Hacker News #1 for two cycles
- Nov 13, 2025
- "Skills Explained" published with progressive-disclosure decision matrices
- Dec 18, 2025
- SKILL.md released as an open standard at agentskills.io
- Jan-Mar 2026
- Marketplace explosion: Antigravity hits 1,400+, SkillsMP indexes 1.2M+
- Mar 2026
- Anthropic Marketplace launches; Skills for Excel and Slides ship
- Apr 2026
- anthropics/skills crosses 117k stars; SKILL.md becomes the de facto cross-vendor standard
§ II S11.2 · The power law of skill stars
Twenty-five repos. One outlier.
Open-source ecosystems are almost always power-law distributed and the Skills ecosystem is no exception. Twenty-five repos account for the lion's share of attention. One of them, obra/superpowers, ranks above Anthropic's own skills repo. That is a meaningful editorial fact: the most-starred Claude Skills project is a community framework, not a first-party release. Hover any bar to see the repo's author class and a one-line note.
About the star counts
Aggregator sites disagree by 30%+ on the same day because some count parent-monorepo stars while others count downstream forks. For consistency, this chart cites GitHub-direct numbers where available and aggregator estimates otherwise. Numbers above 150k (everything-claude-code, superpowers) are directionally accurate but should be verified against the GitHub GraphQL API at a fixed timestamp before any production use.
Note also that anthropics/skills is one repo with
seventeen skills inside; obra/superpowers is one
repo with twenty-plus skills inside. Reporting them as single
bars understates skill-count density. The honest framing is
that two repositories carry the dominant share of community
attention.
- Top community repo
- obra/superpowers, ~178k stars
- Top Anthropic repo
- anthropics/skills, ~117k stars
- Anthropic vs community gap
- community leader is ~52% larger
- Top 25 cumulative
- roughly 1.4M+ aggregated stars
- Star inflation flag
- some 100k+ counts from aggregator scrapers, unverified
§ III S11.3 · The surge timeline
Nine months. Essentially nothing to everything.
The shape of the timeline is the story. Before September 2025, the count is essentially zero. By April 2026, official skills number in the dozens, vendor packs in the dozens, and community skills in the tens of thousands. Three event markers anchor the curve: the public announcement, the open-standard release, and the marketplace explosion.
How the counts are estimated
Anthropic counts come directly from the
anthropics/skills repo's commit history, which lists
seventeen top-level skill folders by April 2026.
Vendor counts are derived from official partner
repositories (Notion, Cloudflare, Stripe, DuckDB, GSAP, Trail
of Bits, HashiCorp, Astral, Sentry, Expo, Microsoft, Google
Workspace, Google Labs, Vercel Labs, Callstack, MiniMax,
Composio, plus a handful of smaller vendors).
Community counts are derived from
SkillsMP's indexed skill claim (1.2M+ by Mar 2026)
cross-checked against sickn33/antigravity-awesome-skills
(1,400+ entries) and the VoltAgent aggregator. The growth curve
between November and February is interpolated; the endpoints
are well-sourced. Treat as directionally accurate, not a
precise time-series.
- Pre-Sep 2025 official count
- 0 (skills internal to Anthropic only)
- Apr 2026 official count
- 17 skills in anthropics/skills
- Apr 2026 vendor count
- ~22 partner packs cataloged
- Apr 2026 community count
- ~100k+ skills aggregated across marketplaces
- Adoption window
- ~6 months from announce to ecosystem maturity
§ IV S11.4 · The architectural sweet spot
Bundles code? Needs an API key?
The dominant architecture is bottom-right: bundles scripts but runs self-contained. That is exactly where Anthropic's official document skills live. Top-right (scripts plus external keys) is where vendor skills congregate. Bottom-left (pure prompt with no external dependencies) is the home of knowledge skills like the Karpathy guidelines. Top-left, scripts you cannot run without external keys plus pure prompt text, is essentially empty. Hover a bubble for details.
Why the empty quadrant matters
Skills are most valuable when they encode both knowledge and capability. Pure-prompt skills (bottom-left) are essentially shared system prompts. Scripts + external creds (top-right) are essentially MCP servers. The bottom-right sweet spot, scripts that run anywhere with no keys, is the Anthropic-recommended canonical form: deterministic actions with no infrastructure burden.
The empty top-left quadrant (pure prompt + external key) is empty for a structural reason: a skill that asks the model to hit an API without giving it any code to do so leaves the model to compose HTTP calls from scratch. That is not value-additive compared to documenting the API in a knowledge skill plus a separate MCP server.
- Sweet spot occupants
- Anthropic 4 doc skills, Superpowers, gstack, Caveman, claude-mem
- Vendor zone occupants
- Stripe, Cloudflare, Notion, Sentry, Composio, plus Anthropic's claude-api
- Knowledge zone occupants
- frontend-design, brand-guidelines, theme-factory, Karpathy guidelines, Humanizer
- Empty quadrant size
- nearly zero (1-2 edge cases out of ~100 cataloged)
- API-key rate (top 25 community)
- ~40% (per awesome-skills.com)
§ V S11.5 · Skills are imperatives
Twelve verbs. Most start with "Use this skill when…"
Skills are not really code. They are imperatives written for an
LLM operator. The opening verbs of a hundred sampled SKILL.md
descriptions converge on a small vocabulary: an imperative
style-marker followed by a content verb. The pattern is so
consistent that Anthropic's own skill-creator skill ships a
script (improve_description.py) that A/B-tests
descriptions against eval queries and grades them on this
exact shape.
Why the descriptions all sound the same
Anthropic's own skill-creator guidance explicitly tells authors
to combat under-triggering by writing imperatively.
"Use this skill when X" is the canonical phrasing because it
doubles as a trigger condition AND a usage note. The
improve_description.py script in skill-creator
automatically rewrites descriptions to match this shape.
The result is a format-wide style convergence. Across vendors (Stripe, Cloudflare, Notion), across community frameworks (Superpowers, gstack), and across knowledge skills (Karpathy, brand-guidelines), the openings are nearly indistinguishable. Skills aren't really documented; they're imperatively trained.
- Sample size
- n = 100 SKILL.md descriptions across Anthropic + vendor + top-25 community
- Style-marker prevalence
- "Use this skill when…" opens 71%
- Top content verbs
- Create (34), Generate (22), Build (19), Extract (17), Design (14)
- Long-tail verbs
- Test, Convert, Analyze, Apply, Manage, Search (each <15)
- Imperative ratio
- ~95% of sampled descriptions begin with an imperative
§ VI Methodology, sources, caveats
Method
Five sections, four hand-coded SVG charts, one editorial frame. The data behind each chart is synthesized from a research brief that itself draws on awesome-skills.com, VoltAgent, Composio's catalog, sickn33's Antigravity Awesome Skills, and direct GitHub star checks for the top twenty-five repos.
Charts are hand-coded SVG with no charting library. Tooltips are vanilla JavaScript, ~120 lines total. Page weight is well under one megabyte. No tracking beyond the platform-level Cloudflare Web Analytics.
Sources
Primary aggregators: awesome-skills.com, VoltAgent, Composio's curated list, sickn33/antigravity-awesome-skills, SkillsMP, agentskills.io (the open-standard home).
Anthropic's own anthropics/skills repo and "Agent Skills" engineering blog are cited where direct numbers were available.
Cross-vendor adoption: tracked through openai/ repos publishing in SKILL.md format, plus Cursor, Gemini CLI, Antigravity, Windsurf, and OpenCode docs.
Caveats
Star counts are point-in-time and inconsistent across aggregators (sometimes off by 30%+ on the same day). Where sources conflict, this lab uses GitHub-direct numbers when available and aggregator estimates otherwise. Numbers above 100k stars (everything-claude-code, superpowers) should be verified directly on GitHub before any production use.
Some highlighted star counts may be inflated by SEO scraper sites; a few popular repos appear inflated by AI-assisted PR farming. Treat 1,000+-skill mega-aggregators as upper bounds, not quality signals.
OpenClaw is a related but separate ecosystem (renamed from Clawdbot due to Anthropic trademark complaint Jan 2026). OpenClaw skills use a similar format but are not directly installable in Claude.
What this lab is not
Not a comprehensive index. ~75 skills are featured here; many more exist. The taxonomy emphasizes breadth (ecosystem shape) over depth (any one skill).
Not a quality ranking. Star counts measure attention, not usefulness. The most-installed skills are Anthropic's four document skills because they ship by default, not because they have the most stars.
Not authored by Anthropic. This is independent editorial analysis. Anthropic press materials are cited where applicable; everything else is synthesis from public aggregators.
Six months ago, "skill" meant nothing in the Claude vocabulary. Today it means a folder with a YAML-fronted Markdown file, and the format is shipping in seven different agent stacks across three vendors. The numbers in this lab will keep moving. The shape of the curve probably will not.
A reasonable working hypothesis: SKILL.md is to LLM agents what
package.json was to Node, what Gemfile
was to Ruby, what requirements.txt was to Python.
A small declarative file that bootstraps a much larger
ecosystem. Boring on its own. Foundational in retrospect.
The next six months will tell us whether the format survives its hype cycle. If you are reading this and the ecosystem now looks twice as big and twice as boring, that is the right shape.
FAQ
What are Claude Skills?
A Claude Skill is a folder containing a SKILL.md file with YAML frontmatter (name, description, optional disable-model-invocation, allowed-tools) plus a Markdown body. Optional sub-folders are scripts, references, and assets. Anthropic announced the format publicly on October 16, 2025 and released SKILL.md as an open standard on December 18, 2025.
How big is the Claude Skills ecosystem?
Six months in: about 17 official skills bundled in anthropics/skills, roughly 22 vendor or partner skill packs from companies including Notion, Cloudflare, Stripe, DuckDB, GSAP, Trail of Bits, HashiCorp, Astral, Sentry, Expo, Microsoft, and Google Labs, plus tens of thousands of community skills aggregated through marketplaces like awesome-skills.com, VoltAgent, Composio, and SkillsMP.
What does the most popular Claude Skill do?
The most-installed individual skills are Anthropic's four document-creation skills (docx, pdf, pptx, xlsx) because they are auto-included in every paid Claude.ai account. The most-starred community framework is obra/superpowers, a 20-plus skill bundle covering test-driven development, brainstorming, planning, parallel agents, and systematic debugging.
Do most Claude Skills require an API key?
No. Of Anthropic's 17 official skills, only claude-api and mcp-builder require an Anthropic API key. Of the top 25 community skills, roughly 60 percent are marked no-credentials. The exceptions tend to be SaaS-bridging or multi-model orchestrator skills like Composio connect-apps, Sentry, Cloudflare, and PAL MCP.
Is SKILL.md becoming a cross-vendor standard?
Yes, early but visible. After Anthropic released SKILL.md as an open standard on December 18, 2025, OpenAI Codex CLI, Cursor, Gemini CLI, Antigravity, Windsurf, and OpenCode adopted it. OpenAI is publishing skills in the same format under openai/security-best-practices, openai/sora, openai/spreadsheet, and others.