wiki

name	wiki
description	Use when the user wants to ingest sources into the wiki, query existing knowledge, or lint the knowledge base for issues. Manages a persistent, interlinked knowledge base (LLM Wiki) with three operations: ingest, query, lint.
tags	["wiki","memory","knowledge","ingest"]

LLM Wiki

A persistent, compounding knowledge base. The LLM writes and maintains the wiki; {{NAME}} curates sources, directs analysis, and asks questions.

Architecture

Three layers:

• Sources (sources/) — raw, immutable input documents (transcripts, articles, PDFs). Never modified by the LLM.
• Wiki (memory/) — flat directory of LLM-generated markdown pages. All pages are siblings. memory/CLAUDE.md is the navigation hub.
• Schema (CLAUDE.md + this skill) — governs structure and operations.

Operations

Ingest

Process a raw source from sources/ into wiki pages. Invoked via the /project:wiki-ingest command.

Input: path to a source file.

Steps:

1. Read the source document fully.
2. Identify key information:
   • Decisions made and their rationale
   • Action items and who owns them
   • People mentioned (names, roles, context) — capture on the relevant project/concept page, never as standalone pages
   • Concepts/themes discussed
   • Contradictions with existing wiki content
3. For each new concept/theme that deserves its own page, create one:

   ---
   type: concept
   name: Concept Name
   first_seen: YYYY-MM-DD
   ---
   

4. Update existing project and concept pages with new information from the source. Add cross-references ([link text](other-page.md)) between related pages.
5. Update memory/CLAUDE.md:
   • Add new pages to the appropriate section (Projects, Concepts, Preferences)
   • Update one-line summaries if they've become stale
6. Delete the source file from sources/ — it has been fully processed.

Rules:

• Never modify files in sources/. They are immutable.
• Never create person pages. People are documented inline on the project or concept pages where they are relevant (e.g., a collaborators table or section).
• One page per entity (project, concept). If a page already exists, update it — don't create a duplicate.
• Every new page must be added to memory/CLAUDE.md.
• Delete each source file after processing — sources/ should be empty when all ingestion is done.
• Use relative links between wiki pages (they're all siblings in memory/).
• Frontmatter type field is required on all pages: project, concept, preference, or synthesis.

Query

Answer questions using the wiki as the primary knowledge source.

Steps:

1. Read memory/CLAUDE.md to identify relevant pages.
2. Read those pages.
3. Synthesize an answer with citations to specific wiki pages.
4. If the answer represents a valuable synthesis (comparison, analysis, connection), offer to file it as a new wiki page of type synthesis.

Rules:

• Always cite which wiki pages informed the answer.
• If the wiki doesn't contain enough information, say so and suggest sources to ingest.

Lint

Health-check the wiki. Run periodically or on request.

Checks:

1. Orphan pages — files in memory/ not listed in memory/CLAUDE.md.
2. Broken links — references to pages that don't exist.
3. Stale content — pages not updated in 30+ days that reference active projects.
4. Missing cross-references — pages that mention entities with their own page but don't link to them.
5. Contradictions — claims in one page that conflict with another.
6. Missing pages — entities mentioned across multiple pages that lack their own dedicated page.
7. Index drift — one-line summaries in memory/CLAUDE.md that no longer match page content.

Output: a report listing issues found and suggested fixes. Apply fixes only with confirmation.