Claude Code skill
The packaged agent workflow — what ships in the repo, what each file is for, and how to use it from any agent.
Marrow ships its agent instructions in the repository, so any agent that reads the working tree gets them for free.
What's in the repo
| File | Purpose |
|---|---|
AGENTS.md | Universal operating router for agents (golden loop, commands, rules, mode router) |
CLAUDE.md | Claude Code project pointer (imports AGENTS.md) |
.agents/skills/marrow-topology/ | Authoring skill — condensed schema, icon guidance, worked examples (mirrored to .claude/skills/) |
.agents/skills/marrow-setup/ | Setup skill — takes a fresh clone to a working state (mirrored to .claude/skills/) |
marrow-topology
Load it when creating or editing topology YAML. It packages the workflow, the condensed schema, and worked examples for each archetype — a small office that lays out tiered, a multi-site campus that lays out zoned, and so on. Copy the closest example and adapt rather than writing from scratch.
The skill defers to AGENTS.md as the authoritative reference; it's the working
checklist, not a second source of truth.
marrow-setup
Load it when the toolchain isn't working: a fresh clone, a missing dist/, or
validate reporting "icon pack is not installed". It installs dependencies,
compiles the toolchain, and fetches the vendor icon packs that are never
committed.
Using Marrow from other agents
Nothing here is Claude-specific. Any coding agent with shell and file access can drive the loop:
npm run marrow -- icons search <keyword> --json # find real icon ids
npm run marrow -- validate input/x.yaml --json # structured diagnostics
npm run marrow -- snapshot input/x.yaml -o /tmp/x.png
npm run marrow -- build input/x.yamlTwo things make the tool agent-friendly:
Structured diagnostics. --json gives code, path, line, column, snippet,
hint, and ranked suggestions. An agent can act on MRW-E010 with a suggestion
list without parsing prose.
Meaningful exit codes. 0 valid, 1 validation errors, 2 usage or I/O
error. An agent can distinguish "the topology is wrong" from "I called the tool
wrong" without reading output at all.
The /api/* endpoints on the dev server exist for the same reason —
GET /api/validate and GET /api/scene let an agent poll state and inspect
computed geometry without scraping HTML. See
Live preview.
Keeping the guides in sync
AGENTS.md stays in the repository because agents read the working tree, not a
website. This documentation site is the canonical, complete reference; the
schema section of AGENTS.md points here rather than restating everything.
If you change the schema, the CLI, or the diagnostic registry, update the
reference pages here — that's the source of truth. Then check whether
AGENTS.md needs a matching one-line change.