LLM Wiki Bootstrap

SkillSearch

Bootstrap and operate an LLM Wiki: a persistent Markdown knowledge base with immutable raw sources, an LLM-maintained wiki, a SCHEMA.md operating contract, thin runtime pointer files, EXTEND.md preferences, and optional local BM25 search. Use when the user asks to create or bootstrap an LLM wiki, ingest sources, query or lint a wiki, configure EXTEND.md preferences, or set up BM25/full-text search for an LLM Wiki.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the LLM Wiki Bootstrap skill

What this skill tells your AI

The instructions your AI receives, as published by nanzhipro/karpathy-llm-wiki-bootstrap-skill in skill/SKILL.md and read by ahel’s review.

Build and maintain a personal LLM Wiki: raw evidence stays immutable in raw/, durable knowledge is compiled into wiki/, and SCHEMA.md is the generated wiki's single operating contract.

This file is intentionally short. Load detailed references only for the operation the user requested.

Operating Model

  • Keep SCHEMA.md as the single source of truth inside every generated wiki.
  • Keep wiki/concept-table.md as the maintained concept map: it complements wiki/index.md by summarizing durable concepts, relationships, sources, status, and maintenance notes.
  • Keep CLAUDE.md, AGENTS.md, and .github/copilot-instructions.md as thin pointers to SCHEMA.md; never duplicate operating rules into them.
  • Treat raw/ as read-only evidence. Create and update derived knowledge only under wiki/.
  • Apply EXTEND.md preferences before bootstrap, ingest, query, lint, or BM25 work.
  • Ask only for missing decisions, and batch setup questions into one user prompt when possible.
  • Prefer progressive disclosure: read the smallest reference file that can complete the current task.

Intent Router

User intentDo thisReference
Create or bootstrap a new wikiRun preference preflight, gather setup answers, then generate the scaffold and seed files.references/workflows/bootstrap.md
Ingest one or more sourcesRun preference preflight, check the optional BM25 gate, then compile source knowledge into wiki/.references/workflows/ingest.md
Answer a domain question from the wikiRun preference preflight, navigate wiki/index.md, use BM25 only as a candidate finder, then cite wiki pages.references/workflows/query.md
Check wiki health or consistencyRun preference preflight, scan structure and content, report findings, then fix only approved items.references/workflows/lint.md
Configure or use BM25 searchLoad preferences, ask before initialization, create local search files, smoke test, and log results.references/workflows/bm25.md
Configure preferencesLocate or create EXTEND.md using the schema and template.references/config/extend-schema.md, references/templates/extend.md

Preference Preflight

Before bootstrap, ingest, query, lint, or BM25 work, read references/config/extend-schema.md and use the first available EXTEND.md in its lookup order. If none exists, run the first-time preference setup and create one from references/templates/extend.md before continuing.

On first use in a session, briefly say which preference file is active. Do not silently use defaults.

Reference Loading Rules

  • For bootstrap, read references/workflows/bootstrap.md, then load each template only when creating that file.
  • For ingest, query, and lint, read the matching workflow file first; read references/workflows/bm25.md only when preferences or the user request require search behavior.
  • For schema generation, read references/templates/schema.md and inject domain snippets from references/templates/domain-page-types.md as needed.
  • For pointer files, read references/templates/agent-pointer.md and keep the generated file minimal.
  • For BM25 initialization, read references/templates/wiki_fts.py and references/templates/bm25-readme.md only after the user approves initialization.

Package Layout Rules

  • Put skill-owned runnable helpers in scripts/ if they are meant to be executed from the installed skill package.
  • Put generated-output templates in references/templates/, even when the generated file will become executable after it is copied into the user's wiki.
  • Put long procedural documentation in references/workflows/ and configuration schemas in references/config/.
  • In this skill, references/templates/wiki_fts.py is a template for the generated wiki's scripts/wiki_fts.py; it is not executed in place by the skill.

Hard Rules

  • Never overwrite an existing wiki without explicit user approval.
  • Never modify files in raw/ except to create empty scaffold placeholders such as .gitkeep during bootstrap.
  • Never answer directly from BM25 snippets; open and read the returned wiki pages first.
  • Preserve source language during ingest unless the user asks for a different wiki language.
  • Update wiki/index.md and append wiki/log.md whenever wiki content changes.
  • If a search index rebuild fails, keep wiki edits, log the failure, and fall back to wiki/index.md plus text search.

Core Outputs

A successful bootstrap creates a wiki root containing:

raw/
wiki/index.md
wiki/concept-table.md
wiki/log.md
wiki/overview.md
SCHEMA.md
.gitignore

Optional outputs depend on setup answers and preferences: runtime pointer files, .vscode/settings.json, scripts/wiki_fts.py, indexes/, and exports/.

Signals

GitHub stars
164
Forks
25
Last commit
May 2026
Advanced
Catalog kind
skill
Gateway key
llm-wiki-bootstrap
Source
github.com/nanzhipro/karpathy-llm-wiki-bootstrap-skill