atlas-skills
Skills that teach an AI agent to model well in Graphite Atlas: the vocabulary, the modeling principles, the order of operations, and the checks.
Also the source for the graphite-atlas Claude Code plugin, which bundles these
skills with the Atlas MCP server — tools plus the know-how to use them. See
README.plugin.md.
The skills
| Skill | What it answers |
|---|---|
atlas-modeling | How should this be modeled? The three principles, each with its anti-patterns inline, and worked examples in atlas-modeling/references/. |
atlas-language | What can I say? Point types, Path types and what may connect to what, properties, the default folders and the naming rules. |
atlas-build | In what order? Reading sources, choosing the unit of work, proposing the change before writing it, labelling what was inferred, writing, and reading it back. |
atlas-auditing | What is wrong? 18 read-only checks that catch principle violations after a build. |
atlas-completeness | What is missing? Absence checks, output as an interview script grouped by who can answer. |
How to use them
- Before any write:
atlas-modeling. - Building from documents (a PDF, an SOP, a transcript, a set of them):
atlas-buildfirst. - Unsure what a type is or what may connect:
atlas-language. - After a build or edit:
atlas-auditing, thenatlas-completeness. In that order: wrong beats thin.
Contributing
Every skill must read as a finished, generic artifact: neutral examples, no people, no internal
references. See CONTRIBUTING.md, and run scripts/check-public.sh before pushing.