A skill is portable expertise for a coding agent. Not a prompt you copy-paste — a small, versioned package the agent loads when the task matches, and ignores when it doesn’t. The registry’s Skill object is modelled on the open Agent Skills format documented in the reference entries.
The package
Every skill ships the same four parts, listed in its files field:
SKILL.md— the entrypoint. Front-matter (name,description) tells the agent when to load it; the body tells it how to do the job.references/— deeper material the agent opens on demand: catalogues, edge cases, decision tables. This keeps the main file short.examples/— concrete, runnable input/output the agent can imitate instead of improvising.tests/— rules or checks that let you verify the skill behaves as documented.
This layering is progressive disclosure: the agent reads the short description first, loads the body only when relevant, and pulls references only when stuck. Context windows are expensive; the format is built to spend them carefully.
What a skill in this registry encodes
Beyond the package, each registry entry declares:
- Learning objectives — what you should be able to do after working through it. Skills double as the hands-on step of the roadmap.
- Prerequisites — other skills to run first, so a complex review doesn’t assume knowledge you don’t have yet.
- Difficulty — beginner / intermediate / advanced.
- Compatibility — the agent runtimes the skill has been written for (e.g.
claude-code,codex,cursor,amp,gemini-cli). - Security flags — whether the skill touches the filesystem, runs code, calls the network, or needs credentials (see Security).
- Provenance — license, author, source repo, and the install command.
A minimal example
The registry’s own skills follow this shape — for instance a dbt skill lists SKILL.md, references/incremental-patterns.md, examples/models.yml and tests/test_rules.py, with prerequisites pointing at a more basic SQL skill. Skills compose: review skills assume the SQL review skill, streaming skills assume orchestration basics.
Writing a good one
- Name the trigger. The
descriptionis what the agent matches against — “reviews a dbt model for grain and incremental correctness” beats “dbt helper”. - Rank by blast radius. Put the failure that costs the most first (e.g. grain tests before stylistic ones).
- Show, don’t tell. One worked example beats three paragraphs of advice.
- Declare your blast radius. If the skill never runs code, say so — the security flags let users filter on exactly that.
New skills enter the registry as curated YAML records; see Contributing and the content object model.