knowledge-mcp: Skill for Claude Code

Ingest patterns: $ARGUMENTS

For universal orchestration discipline reference /orchestrate. This skill is pattern-ingestion-specific.

The flow is: collect, find, read, land, polish, drop. Step 0 orients, and steps 1 through 7 run in that order.

Reading a collected source is cheap and repeatable — the raw graph is a cached structured document, and every read after the collect replays it and fetches nothing over the network. So the deciding happens while reading. The writing at the end is hand work on the small set you already chose.

Step 0: Orient

Check the server is up and see what is already collected.

manage({ "operation": "status" })

query({ "graph": "web", "mode": "modules" })

query({ "graph": "pdf", "mode": "modules" })

mode:"modules" lists the raw graphs already collected, with their node counts. Run it for whichever family your source belongs to, or both if you are not sure which one holds it.

Pick a SCRATCH NAME for this run — something you will not mind losing, keyed to the source and this session. Step 7 drops the graph.

Step 1: Collect the source into a raw graph

Web:

collect({
  "type": "web",
  "id": "<scratch-slug>",
  "seed_urls": ["<entry-url>"],
  "follow_patterns": ["^<host-and-path-prefix-regex>"],
  "max_pages": 100,
  "politeness_ms": 500
})

PDF:

collect({ "type": "pdf", "id": "<absolute-path-to-pdf>" })

max_pages is a HARD cap on the pages the crawl fetches, not a target it approaches — the crawl stops there even with work still queued, and zero means unbounded. follow_patterns is a regex allowlist for internal links; anchor it at the host and path prefix you actually want, or a site-wide crawl will bring back mostly navigation.

Pass no force on either call. A scratch name has nothing to overwrite, and force is refused outright on a recipe run later in this flow.

Then read the shape of what you collected, before writing any body: