| name | wiki |
| description | Build and maintain the agent wiki (.claude/wiki/) — a persistent, compounding knowledge base. Use when ingesting a source into the wiki, querying it, or linting it for drift. The wiki links to the SSOT (code / _docs/ / _note/); it routes and synthesizes, it does NOT duplicate facts. |
Wiki
The agent's compounding knowledge layer. Unlike RAG (re-derive from raw docs on every query), the wiki is built once and kept current — an interlinked set of markdown pages under .claude/wiki/ that the agent owns and maintains.
Three layers (Karpathy LLM Wiki): raw sources (_note/ human, _docs/ project, learnings/, external) → the wiki (.claude/wiki/, this skill owns) → the schema (.claude/wiki/schema.md, conventions).
Where it lives
.claude/wiki/ is the agent bucket (see docs-lifecycle Three-bucket section). Bootstrapped by /team-init:
index.md — content catalog (entities / concepts / sources). Read FIRST on query.log.md — append-only chronicle, prefix ## [YYYY-MM-DD] ingest|query|lint | <title>.schema.md — structure + conventions (read before any operation).<page>.md — entity / concept / overview / comparison pages.
Core rule: link, don't duplicate. A page routes to and synthesizes the SSOT (code, _docs/, _note/); it never restates a fact that lives elsewhere. A duplicated fact goes stale. (continuous-learning §7.)
Operation: ingest
User-driven, one source at a time, supervised (the preferred flow). Trigger: the human points at a source ("ingest this").
- Read
schema.md for current conventions.
- Read the source. Extract the key information (do not copy it verbatim).
- Discuss takeaways with the human if non-trivial.
- Write/update the summary page for the source.
- Update entity/concept pages the source touches — a single source may touch 10–15 pages.
- Add cross-references between affected pages; link to the SSOT (code /
_docs/ / _note/).
- Update
index.md (new/changed pages, one-line summaries).
- Append to
log.md: ## [<date>] ingest | <source title>.
Sources: _note/ (read-only — read for context, never modify), _docs/ (on completion), .claude/session-state/learnings/, external docs the user provides.
Operation: query
- Read
index.md to find relevant pages.
- Drill into those pages; follow cross-links.
- Answer with citations (the page + the SSOT it links to).
- File good answers back. A useful comparison, analysis, or discovered connection is a new wiki page — do not let it vanish into chat. Update
index.md + append ## [<date>] query | <question> to log.md.
Operation: lint
Periodic health check (when a page feels stale, at workflow completion, or on request). This IS the continuous-learning §7 self-audit applied to the wiki — reuse it, do not reinvent.
Check for:
- Contradictions between pages.
- Stale claims newer sources/code have superseded — follow a page's claims to the code and flag any that no longer hold (verify it still exists).
- Orphans — pages with no inbound links.
- Missing cross-references between related pages.
- Concept gaps — important concepts mentioned but lacking their own page; data gaps a search could fill.
Output: fixes applied + suggested new questions/sources to investigate. Append ## [<date>] lint | <scope> to log.md.
Relationship to continuous-learning (no overlap)
- continuous-learning owns the lifecycle of reusable patterns/instincts (HOW to work) — learnings, confidence scoring, skill evolution.
- wiki owns the synthesis of knowledge/facts (WHAT is true) about a domain/codebase.
learnings/ is one ingest source for the wiki. A high-confidence, project-stable learning MAY be promoted to a wiki page (in parallel with §6 profile promotion — routing, not duplication).- continuous-learning §7 governs wiki maintenance (link-don't-duplicate, same-change-same-update, self-audit = lint).
Scale note
At small scale the index.md catalog is enough to navigate — no embedding/RAG infrastructure needed. If the wiki outgrows comfortable index navigation, a local markdown search tool (e.g. qmd) is a future option, not a requirement.
