/llm-wiki

Summary

Set up or operate a general-domain LLM Wiki (Karpathy pattern) — a persistent, LLM-maintained markdown knowledge base for research, personal knowledge management, book reading, course notes, competitive analysis, or any non-product domain.

When to use

Use /llm-wiki for any general-domain knowledge management need that is not tied to building a product. It applies the Karpathy wiki pattern with a flat raw/ directory and page categories of wiki/topics/ and wiki/decisions/.

Trigger phrases (from SKILL.md):

• "set up an LLM wiki"
• "apply the Karpathy wiki pattern"
• "build a persistent knowledge base"
• "ingest into the wiki"
• "create a topic page"
• "draft a document from the wiki"
• "lint the wiki"
• Any task involving the llm-wiki/ directory with raw/, wiki/, drafts/ layers

Invocation

Invoke as a slash command. The skill routes to Setup, Operate, or Repair automatically based on the state of llm-wiki/:

/llm-wiki

The skill determines which mode to run:

• No llm-wiki/ directory exists → runs Setup (Step 1): creates the full directory tree, copies templates, adapts schema.md to the project, and offers pointer injection into CLAUDE.md/AGENTS.md.
• llm-wiki/ exists and is complete (contains schema.md and full layer tree) → runs the requested Operation (Step 2): ingest, create topic page, draft, query, lint, archive, evolve schema, or watcher commands.
• llm-wiki/ exists but is incomplete → runs Repair (Step 1b): fills in missing pieces without overwriting existing files.

Operations available once the wiki is set up:

• ingest — read a source, update wiki pages, update index.md, append to log.md
• create a topic page — synthesize a wiki/topics/<name>.md
• draft a document — write a drafts/<name>.md
• query — read index.md, find pages, synthesize answer with citations
• lint — health-check the wiki (orphans, missing citations, contradictions)
• watcher on / watcher off — enable/disable and start/stop the filesystem watcher for raw/ (desired state persists across restarts)
• watcher — report watcher liveness plus the files that need ingesting, and offer to ingest them. Ingest state is owned by a script ledger (catalog.py) that diffs raw/ into new / changed / current / missing; the watcher is opt-in observability only.

Example 1 — Setting up a new wiki for a research project

/llm-wiki
Set up an LLM wiki for my machine learning research project.

No llm-wiki/ directory found. Running Setup (Step 1).

Created:
  llm-wiki/
    karpathy-llm-wiki.md   ← reference copy
    schema.md              ← adapted to project (docs/ and papers/ detected as external roots)
    index.md
    log.md
    raw/
    wiki/
      overview.md
      glossary.md
      topics/
      decisions/
    drafts/

Detected external documentation roots:
  | Path      | Note                                                     |
  |-----------|----------------------------------------------------------|
  | docs/     | Read-only literature corpus — cite as docs/<f>.          |
  | papers/   | Research papers — read for context, never modify.        |

Log entry appended: [2026-06-05] init | wiki created | schema.md, index.md, wiki/overview.md, wiki/glossary.md

Offer: Inject llm-wiki pointer into CLAUDE.md? (yes/no)

Example 2 — Ingesting a source into an existing wiki

/llm-wiki
Ingest this paper abstract into the wiki as a topic page on "attention mechanisms":
[paste of paper abstract on transformer attention]

Running Operation: Ingest

1. Saved raw source to: llm-wiki/raw/attention-mechanisms-vaswani-2017.md
   (Provenance header added: source = "pasted abstract", date = 2026-06-05)

2. Updated wiki page: llm-wiki/wiki/topics/attention-mechanisms.md
   - Added section: "Scaled dot-product attention (Vaswani et al., 2017)"
   - No conflicts detected with existing content.

3. Updated index.md: added entry for "attention mechanisms"

4. Appended to log.md:
   ## [2026-06-05] ingest | attention-mechanisms-vaswani-2017.md | wiki/topics/attention-mechanisms.md

Next recommended operation: query or lint to verify the wiki state.