wikilisted
Install: claude install-skill nathanaday/claude-atlas
# Orientation
Atlas has two things. A **knowledge base** is an Obsidian vault: `inbox/`
for sources you have not processed, `ideas/` for the user's own scratch
notes, `.raw/captured/` for immutable copies of ingested sources, and
`wiki/` for the pages: sources, entities, and concepts. Every change to
`wiki/` is one reviewed operation and one git commit. A **project** is a
folder `atlas/<name>/` inside the user's work, a repository or a folder of
documents. It holds `project.json`, the threads (`threads/`, and one folder per
stage: `stubs/`, `specs/`, `plans/`, `receipts/`), `phases/`, and `inbox/`
for notes. It has no git of its own and no engine. A project uses one knowledge
base; a knowledge base serves many projects.
The atlas MCP server (tools named `mcp__plugin_claude-atlas_atlas__<tool>`,
called `status`, `plan`, `apply`, and so on below) is the only write path
into a knowledge base.
## Find the place
Call `status` first. In a project session it reports `kind: project`, the
project's id, name, description, and path, its knowledge base (or why none
resolves), the page that describes it there, and its thread counts by stage. In a
knowledge base session it reports `kind: knowledge`, the mode, scope, page
count, files waiting in the inbox, git state, warnings, and the projects that
use it. The session hook's first line already names the place:
`claude-atlas: project …` or `claude-atlas: knowledge base …`.
If `status` fails because the session is in neither, hand off: `