knowledge-mcp: Skill for Claude Code

WRITE-A-BRIEF — everything load-bearing goes in before the spawn

A brief is a decision, not a relay

A brief is where the orchestrator's reading of the situation becomes a lane's instructions. It is written after the return protocol, never from a report's summary line. If you cannot say in one sentence what this brief is for and what it demands that the situation requires, you are not ready to write it.

Discovery, never confirmation

A research brief names the problem and asks what already exists; it never names a solution. An investigation brief carries the observation (what ran, what it returned, on which tree and plane) and the instruments; it never carries a candidate mechanism. A planning brief names the ticket and the tree; it never names a design. If your brief contains a mechanism you have not reproduced or a design the user has not decided, you have poisoned the lane.

Front-load everything

A lane reads its mailbox only between turns and consumes queued messages after its current work, so a mid-flight correction is a note found later, maybe never acted on, and a message to an idle lane resumes it. Everything load-bearing goes into the spawn prompt. A post-spawn correction is a re-spawn, or a follow-up sent only after the lane's idle notice with the lane then treated as live again.

The brief that opens a round states the sha the round is measured on and is the only message that may instruct a rebase. A brief that opens a fix round says: the report that declares the round done follows the round's one commit and is the hand-off; an unfinished item is stated in that report, never committed after it; the implementer never pushes, the orchestrator does.

Standing lines in every brief

• The artifact ids (ticket, prefill, commit) and the tree and branch to work at.
• Every finding, annotation, decision and ticket id in a brief is copied in full from a graph read you made for this brief, never transcribed from a report's prose. A prefix or a relayed id sends an auditor to resolve a mismatch instead of auditing.
• The name of a landed artifact (a manifest, a contract file, a schema) is relayed only after you opened it on the landed tree; a sibling lane's report of its spelling is not the artifact.
• ABSOLUTE PATHS for every read, build and run; a persistent shell cwd drifts.
• The user's load-bearing rules (money, access, security, data handling, scope) as VERBATIM QUOTES, with your interpretation beside the quote, labeled yours.
• The isolation rule: tests and probes run against spawned services on picked ports with an isolated home; the operator's running services, stores and credentials are never restarted, reconfigured or written into.
• The reads to stamp, always including .claude/skills/knowledge-tools/SKILL.md, and the line that the report ends with a tool census.
• Sibling work in flight on the same branch, with its touched files.
• The repository's own standing test and commit rules, quoted from where the repository states them.
• Any single operation projected over fifteen minutes is named before it runs.
• The report's size bound and the order of its contents, with the verdict, the ids and the not-done list first, so a report truncated in transit still carries what the next decision needs.