Documentation
The format, the gates, and the surfaces.
The skill format
A skill is a directory containing SKILL.md — a Markdown file with YAML frontmatter carrying a `name` and a `description`. The description is what an agent reads to decide whether to load the skill, so it must state both what the skill does and when to use it. Supporting files live alongside it and are referenced from the body, so the agent only reads them when it needs them.
Quality gates
Gate 0 validates the frontmatter and the schema. Gate 1 lints the structure — progressive disclosure, size limits, no absolute paths, no leaked secrets. Both run inside the request, in under a fifth of a second. Gate 2 runs a trigger evaluation asynchronously: a judge model checks whether the skill fires on prompts it should, and stays silent on prompts it should not. Gate 3 executes any bundled scripts in a sandbox.
Installing a skill
Every skill page shows a one-line install command. It writes the skill directory into your project so your coding agent picks it up on the next run. Nothing is installed globally and nothing runs at install time.
Agent-readable surfaces
The registry publishes itself in the formats agents look for: /llms.txt as a Markdown index, /index.json as a typed index, and a clean Markdown rendering of every skill. A Content-Signal header declares how the content may be used. This is the same machinery the skills themselves are about.
Example SKILL.md
---
name: core-web-vitals-audit
description: Audits a web application for Core Web Vitals
regressions. Use when the user asks to check LCP, INP or
CLS, or investigate a Lighthouse score drop.
---
# Core Web Vitals audit
## When to use this
...
## Procedure
1. Measure first, guess never.
2. See [references/thresholds.md](references/thresholds.md).