Atlas Skills & Plugin
Connecting the MCP server gives your AI hands — tools to read and write your Atlas. It doesn't give it judgment. Left to itself, an AI will build a messy model: vendor names baked into point names, dependencies hung off the wrong level, notifications typed as handoffs.
The Atlas skills fix that. They're a set of instructions your AI reads before it touches your Atlas, so it models the way an experienced Atlas user would.
The skills
| Skill | What it does |
|---|---|
atlas-modeling | The three modeling principles, each with its common mistakes listed inline. Your AI reads this before writing anything. |
atlas-language | The Atlas language: which point types and path types exist, what can legally connect to what, and where things belong. |
atlas-auditing | Read-only health checks that catch modeling problems after a build. The backstop for whatever prevention missed. |
The three principles they teach:
- Name by function. A point is named for what it functionally is. Vendor, person, and per-flow context live on paths and properties, not baked into the name.
- Attach at the right level and type. Dependencies hang off the atomic step. Services are Systems connected to Vendors, never Step → Vendor directly.
- Type by meaning, membership explicit. Sequenced steps each need their own membership path. A handoff is a transfer of responsibility, not a notification.
Install the plugin (recommended)
The plugin bundles the skills and the MCP server together, so one install gets you tools plus the know-how to use them well.
In Claude Code:
/plugin marketplace add graphiteatlas/atlas-skills
/plugin install graphite-atlas
Then set your API key in your environment:
export GRAPHITE_ACCESS_TOKEN=your-token-here
Restart Claude Code. You'll have the Atlas tools and the skills loaded together, plus two commands: /atlas-model to model something, /atlas-audit to health-check what you built.
Use the skills on their own
If you've already connected the MCP server and just want the modeling know-how, the skills work standalone. Clone the repo and point your AI at it:
git clone https://github.com/graphiteatlas/atlas-skills.git
In Claude Code, symlink the skill folders into ~/.claude/skills/ and they'll be discovered across all your sessions. In other tools, point your AI at the repo and ask it to read skills/atlas-modeling/SKILL.md before building.
If you installed the plugin, don't also symlink the skills. They'd load twice.
Tools without a skills mechanism
Most AI platforms have no equivalent to skills. Copilot Studio, for instance, gives an agent your tools but no way to load instructions that always apply. For those, use the starter files:
- Download agent instructions → — for the agent's instructions or system prompt field. This is the part that's always in context, so the sharpest rules live here.
- Download knowledge file → — for platforms that accept an uploaded reference document.
They're a condensed stand-in, not a replacement. A skill loads deterministically right before the model acts; an uploaded document is retrieved only when the platform decides it's relevant. Expect a competent read-only agent from the starter files, and keep authoring where the skills actually load.
Verify it works
Ask your AI to model something small: "Model our invoice approval process in Atlas."
A model with the skills loaded will invoke atlas-modeling first, ask what it needs to know, then build with correct types and attachment. Without them, it starts creating points immediately and you'll see the mistakes above.
Afterward, ask it to "audit the atlas" and it should run the health checks and report what it found.
Contributing
The repo is public and takes contributions. New modeling patterns get added as examples under an existing principle rather than as new skills. CONTRIBUTING.md has the authoring rules and a skill template.
github.com/graphiteatlas/atlas-skills →
Questions? Email us at [email protected]