From woz
Inspects and manages the WozCode knowledge base layers (company, repo, personal notes). Use for status, search, notes, suppression, boosting, or ingest.
How this skill is triggered — by the user, by Claude, or both
Slash command
/woz:woz-knowledgeThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
The WozCode knowledge base has three layers, queried bottom-up:
The WozCode knowledge base has three layers, queried bottom-up:
| Layer | Source | Sync? |
|---|---|---|
| company | org-wide chunks (cross-repo learned rules) | server (v2) |
| repo | code files + PR history + distilled rules | local (v1) / server (v2) |
| personal | your notes, suppressions, boosts | local (v1) / server (v2) |
This skill is a thin wrapper around the wozcode knowledge CLI subcommand.
It's the right tool when the user wants to:
status)query)note)suppress)boost)ingest)refresh)ops)TRIGGER on:
/woz-recall which searches past Claude Code sessions)/woz-knowledge literallyDO NOT use this skill for:
mcp__plugin_woz_code__Search is the right tool there./woz-recall searches Claude Code session transcripts, not the knowledge base./woz-review.Map the user's intent to one of the subcommands. When ambiguous, prefer status + a clarifying question over a guess.
| User says... | Subcommand |
|---|---|
| status / what's indexed / health check | status |
| search / find / look for | query <text> |
| remember / note / always use | note "<text>" (add --repo only when user explicitly says "for this repo") |
| remove note / forget note | unnote <noteId> (add --repo if the note is repo-scoped) |
| suppress / hide / stop showing | suppress <chunkId> |
| undo suppress / unhide | unsuppress <chunkId> |
| boost / weight higher | boost <chunkId> <factor> |
| undo boost | unboost <chunkId> |
| ingest / add this file | ingest <path> |
| refresh / rebuild | refresh |
| my notes / show overlay | ops |
For suppress / unsuppress / boost / unboost, the user almost
never knows the chunkId off the top of their head. The natural flow is:
query <their phrasing> first to surface candidate hits with ids.Run WITHOUT 2>/dev/null — stderr surfaces useful errors (login required,
not in a github repo, etc.).
node --no-warnings=ExperimentalWarning ${CLAUDE_PLUGIN_ROOT}/scripts/woz-knowledge.js <subcommand> [args] [--json]
Use --json when you need to programmatically inspect output (e.g. to
follow up with another subcommand). Use the human-readable output
when you'll just print it back to the user verbatim.
For status and query: print the CLI output verbatim. Don't
paraphrase the layers / hits — the formatting is already user-facing.
For note / unnote / suppress / boost / unsuppress / unboost: print
the one-line confirmation the CLI emits, followed by the new op id so
the user can undo if they want.
For ingest: print the chunks-added summary.
For refresh: print the jobId. The local provider's refresh is
fire-and-forget; results land in the next status call.
For ops: if the user is reviewing what they've done, print verbatim.
If they're looking for something specific ("what did I suppress yesterday?"),
filter the JSON output and present a focused list.
Personal-overlay subcommands (note, suppress, boost, etc.) require
login. The CLI surfaces "login required" on stderr; if you see that,
tell the user to run /woz-login and retry. Do NOT silently no-op.
github:owner/repo). If the
user is in a non-GitHub repo or a worktree without origin, repo-scoped
ops will skip the repo layer and only operate on personal-global notes.'remote' by default) is an internal setting — it is not configurable
via /woz-settings. Same CLI surface either way.query runs against every layer the user has access to and merges
results with overlay ops applied. The [scope/kind] tag on each hit
shows where it came from.npx claudepluginhub withwoz/wozcode-plugin --plugin wozLeverages fireauto-mem DB to actively search and apply user's accumulated dev knowledge on files, topics, patterns during code mods, bug fixes, new features, or explicit history requests.
Captures PR review findings into project memory palaces as decisions, patterns, standards, and lessons. Retrieves for reference in future reviews and code work.