The judgment steps, and the skills that carry them

Three steps need a reader: deciding what the corpus must answer, deciding the vocabulary, and stating which facts each document asserts. The code makes those steps safe rather than automatic.

The playbooks under skills/ run them with an agent, and stop for a person at the gates. build-knowledge-base is the end-to-end path; the others are its steps: the interview, curation, provenance vetting, querying, evaluation, capture, which is how a conversation becomes a source, and spec, the downstream use: a product spec or decision record drawn from the graph with every claim cited, whose decisions go back into it as DecisionRecord facts under a convention.

Beside them, concierge does no task: it reads oto status, explains the station the project is at, offers the options at the fork in front of the person, prints the recorded reason behind a class, a rule or a change (oto ontology rationale, oto rules explain, the ledger), and names the playbook to load.

It quotes commands rather than restating them, and a test fails the build when it names a command the engine lacks or when oto status can suggest a step the skill does not know. An ontology may carry a guide.md, what its domain's graph is for and how to read it, installed as GUIDE.md and quoted by the concierge. act is the protocol for the graph's hands: list the actions the graph declares, confirm by name before anything that changes the world, invoke through the declared transport with the agent's own tools, record the response as evidence.

The .claude-plugin/ manifest and the .claude/skills/ links make them discoverable by Claude Code; .mcp.json registers the query server for the open project and hooks/hooks.json reports the project's state at session start.