llm-wiki-setup

Install the Karpathy LLM wiki pattern into a folder. One pass, three questions, a working vault a beginner can use immediately.

What gets created

<vault>/
├── CLAUDE.md              # the map — routing table, schema, workflows, guardrails
├── AGENTS.md              # minimal mirror for Codex and other non-Claude agents
├── .gitignore
├── raw/                   # immutable source documents (LLM reads, never writes)
│   └── .gitkeep
└── wiki/                  # LLM-owned markdown
    ├── index.md           # catalog of every page, organized by category
    └── log.md             # chronological operation log with grep-friendly prefix

Optional additions the user may request at setup time:

• wiki/hot.md — 500-char rolling buffer for fast recent-context lookup (executive-assistant use cases)
• Nested wiki subfolders: wiki/entities/, wiki/concepts/, wiki/sources/, wiki/analyses/

Setup flow

Ask these three questions in a single message:

1. Vault path? Absolute path. Will be created if it doesn't exist.
2. Flat or nested wiki? Flat is the default (Karpathy's preference — cleaner, grep-friendly). Nested adds subfolders for larger or mixed-topic vaults.
3. Include hot cache? Optional wiki/hot.md rolling buffer. Default: no.

Build procedure

1. Safety check. If <vault>/CLAUDE.md or <vault>/wiki/ already exists, stop and ask the user — do not overwrite without explicit confirmation.
2. Create directories. <vault>, <vault>/raw/, <vault>/wiki/. For nested layout, also create wiki/entities/, wiki/concepts/, wiki/sources/, wiki/analyses/.
3. Install the schema. Copy templates/CLAUDE.md → <vault>/CLAUDE.md, replacing {{LAYOUT}} with flat or nested. Copy templates/AGENTS.md → <vault>/AGENTS.md.
4. Scaffold the wiki. Copy templates/index.md → <vault>/wiki/index.md. Copy templates/log.md → <vault>/wiki/log.md, replacing {{DATE}} with today's date (YYYY-MM-DD) and {{LAYOUT}} with the layout choice.
5. Optional extras. If hot cache was requested, create <vault>/wiki/hot.md with a header comment only. Create <vault>/raw/.gitkeep.
6. Gitignore. If <vault>/.gitignore does not exist, copy templates/gitignore → <vault>/.gitignore. If it already exists, append any missing lines from the template — do not overwrite.
7. Git init. If <vault> is not already a git repo: git init && git add . && git commit -m "chore: initialize LLM wiki". If it is already a git repo with a clean working tree, stage the new files and commit them separately as chore: add LLM wiki scaffold. If the repo has uncommitted changes, stop and ask the user how to proceed.
8. Verify. Read <vault>/CLAUDE.md back and confirm the {{LAYOUT}} placeholder was replaced. If it wasn't, fix it and amend (or create a follow-up commit).

Report to user

On completion, print exactly:

LLM wiki initialized at <vault>

Layout: <flat|nested>
Hot cache: <yes|no>

Next steps:
1. Drop a source into <vault>/raw/ (PDF, markdown, web clip, transcript)
2. Open Claude Code in the vault:  cd <vault> && claude
3. Say: "Ingest the new source I just added to raw/"

Claude will read it, extract entities and concepts, write wiki pages,
update wiki/index.md, and append to wiki/log.md. A substantive source
typically touches 10–15 pages — that is normal and the whole point.

For periodic gap-filling research, install and run the wiki-self-heal skill.

Notes

• The CLAUDE.md template is the hero file. It encodes the three-layer architecture, the routing table, naming conventions, workflows, and guardrails. Beginners read it to understand the system — it's also documentation, not just instructions for the LLM. Keep it intact during the copy; only {{LAYOUT}} is substituted.
• The AGENTS.md template is a minimal mirror for OpenAI Codex and other non-Claude agents that don't read CLAUDE.md.
• Never overwrite existing files without explicit user confirmation (steps 1 and 6).
• The vault is just a git repo of markdown files. No database, no embeddings, no vector store. That is intentional — naming conventions and wiki/index.md replace RAG infrastructure at this scale.