Use when an existing llm-wiki has outgrown its flat structure — symptoms include silent contradictions piling up, index.md too long for one read (>500 lines), stale pages accumulating, users asking relationship questions flat wikilinks can't answer ("show me everything that touches X"), users forgetting to run lint for weeks, multiple agents sharing the same wiki, or sensitive sources that need governance. Not for fresh projects or wikis under ~100 pages — start with llm-wiki instead.
REQUIRED BACKGROUND: You MUST use llm-wiki first to bootstrap and understand the minimal pattern. This skill extends it; it does not replace it. If ./wiki/ and ./wiki/SCHEMA.md do not exist yet, stop and run llm-wiki first. Everything in this skill assumes the minimal pattern is already in place.
This skill adds layers. It does not change the three-layer architecture (raw/ → wiki/ → wiki/SCHEMA.md), the canonical filenames, the four page types, the ingest/query/lint operations, or the log prefix format. It adds new frontmatter fields, new operations (crystallize, forget), and new infrastructure (hybrid search, automation hooks).
Knowledge has a lifecycle. Structure earns its keep.
A flat wiki compounds beautifully up to roughly 100 pages. Past that, three things go wrong:
index.md grows past the point where reading the whole thing per query is efficient, and keyword wikilinks can't express "everything that depends on X."This skill adds the machinery that handles all three: confidence and supersession for trust, typed edges and hybrid retrieval for queries, consolidation tiers and automation for maintenance. Add layers as you feel the pain. Not before.
You do not apply this skill as a bundle. You apply specific layers when specific symptoms appear. Use this table to pick.
| Symptom the user describes | Layer to add | Section below |
|---|---|---|
| "I found contradictions sitting silently next to each other" | Lifecycle — confidence + supersession | Lifecycle Layer |
| "The wiki feels cluttered with old stuff I don't want to delete" | Lifecycle — tiers + forget operation | Lifecycle Layer |
"index.md is too long, queries miss pages" | Break the index, add hybrid retrieval | Retrieval Layer |
| "Show me everything that touches X (relationship query)" | Knowledge graph with typed edges | Graph Layer |
| "I keep forgetting to run lint" | Automation hooks + scheduled agents | Automation Layer |
| "Multiple agents or users sharing the wiki" | Shared/private scoping + mesh sync | Collaboration Layer |
| "Sources contain API keys or PII" | Privacy filter on ingest + audit trail | Privacy Layer |
| "Completed research threads should become wiki pages" | Crystallization operation | Crystallization |
Addressing symptoms in this order works because later layers depend on frontmatter introduced by earlier ones. Always do Lifecycle before Graph before Retrieval — the graph depends on edge frontmatter, and hybrid retrieval wants access to confidence and tier for ranking.
Adds trust scoring and structured retirement to every page.
---