Your AI Maintains the Wiki, You Just Feed It

正體中文

You know that knowledge base you "started organizing properly" three times last year? Yeah. This one's different — because the one maintaining it isn't you.

Production-ready template based on Andrej Karpathy's 2026-04 LLM Wiki gist. Three-tier architecture (raw / wiki / schema) plus a journal layer extension. Clone it, feed it docs, let the LLM do the bookkeeping humans always abandon.

Why not just use Obsidian / Notion / Confluence

Those tools assume you maintain the knowledge base. Cross-references, summaries, contradictions, missing pages — somebody has to keep it consistent. Humans give up around page 50.

LLMs don't give up. This template is the structure that lets them help.

• LLM maintenance cost ≈ zero: no fatigue from updating cross-refs, no skipping bookkeeping
• Three tiers, clean separation: raw = source of truth, wiki = LLM's organized output, schema = the rules
• Git native: version control, diff, optional git-crypt encryption
• No vendor lock-in: pure markdown — switch LLMs or tools tomorrow, take your wiki with you

Quick Start

git clone https://github.com/YOUR/kc_llm_wiki_starter.git my-wiki
cd my-wiki

1. Replace {{ your-wiki-name }} placeholders in CLAUDE.md / README.md / index.md / log.md
2. Read wiki/_example_entity.md and wiki/_example_overview.md to learn the frontmatter convention
3. Delete examples (git rm wiki/_example_*.md) or keep them as reference
4. Open a Claude session, drop a doc into raw/, and say "ingest this into wiki"

Full walkthrough: docs/getting-started.md

Core Concepts (30-second version)

raw/     ← External dumps (gitignored, not in git)
wiki/    ← LLM-maintained topic pages (frontmatter + body)
journal/ ← Temporal work logs (daily / meeting / checklist)
SCHEMA.md, log.md, index.md  ← meta layer

Three workflows:

1. Ingest: feed raw doc → agent confirms key points with you → writes to wiki → updates index + log
2. Query: ask agent → reads index → reads wiki → answers with citations
3. Lint: run /llm-wiki-lint periodically → catches stale claims, orphan pages, missing topics, data gaps

Detailed: docs/workflow.md

Companion skill

kc_ai_skills/llm-wiki-lint — because once your wiki passes ~15 pages, stale claims, broken cross-refs, and missing topics start rotting silently. Humans don't catch this. The lint skill does.

This template's CLAUDE.md already includes a session-start hygiene check — your agent will remind you to run lint when it's been more than 14 days since the last one.

Install:

git clone https://github.com/KerberosClaw/kc_ai_skills.git
cp -r kc_ai_skills/llm-wiki-lint ~/.claude/skills/

Security Notice

This template ships with no actual data, credentials, or external dependencies — it's just markdown scaffolding. That said:

• Your raw/ folder is gitignored by design. But anything you put there lives on your local disk — plan your own backup / sync strategy.
• Commit messages are NOT encrypted by git-crypt, even if you enable it. Keep sensitive entity names out of commit messages.
• External LLM APIs: when your agent ingests / queries, content is sent to whichever LLM you use (Claude, GPT, etc.). Apply your own data classification policy before feeding sensitive material.
• Report security issues: open a GitHub issue (this template repo is intended for non-sensitive content).

Documentation

• docs/getting-started.md — 5-minute first ingest
• docs/workflow.md — Ingest / Query / Lint deep dive
• docs/extensions.md — Customization (journal layer, frontmatter extensions, tool integrations)

License

MIT — see LICENSE.