llm-wiki

LLM Wiki

A persistent, compounding personal knowledge base. The wiki grows richer with every source you ingest and every question you answer. Unlike RAG, knowledge is compiled once and kept current — not re-derived on every query.

Based on Karpathy's LLM Wiki pattern: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Directory Structure

All wiki files live in wiki/ inside the workspace ($WIKI_DIR, defaults to ./wiki/ if not set):

wiki/
├── index.md        ← catalog of all pages (update on every ingest)
├── log.md          ← append-only chronological record of operations
├── raw/            ← source documents (immutable — read only, never modify)
└── pages/          ← LLM-maintained wiki pages (you own this layer entirely)

On first use, create this structure if it doesn't exist.

$SKILL_DIR in script commands means this skill's installed directory. If your agent does not set it automatically, replace it with the path to this llm-wiki directory before running commands.

Page Conventions

Every page in wiki/pages/ should follow this format:

---
title: Page Title
tags: [tag1, tag2]
sources: 0
last-updated: YYYY-MM-DD
---

# Page Title

[body content]

## See also

- [Related Page](related-page.md)

• Filename: lowercase slug, e.g. machine-learning-basics.md
• Always include a ## See also section with cross-links to related pages
• Update sources: count when a new source touches this page
• Update last-updated: on every edit

Operations

List the wiki

When the user asks to list, show, or browse their wiki — e.g. "list your wikis", "list my wiki", "list wiki pages", "what's in my wiki", "show my notes":

1. Run: python3 "$SKILL_DIR/scripts/list.py" (add --tag TAG or --sort updated if the user narrows it)
2. Present the returned pages as a readable table (title, tags, sources, last-updated).
3. If the script returns an empty array [], the wiki has no pages yet — say so and offer to start one (do not describe the directory layout as if it were the answer).

Do not answer this request by paraphrasing the Directory Structure section — always run list.py and report the actual pages.

Ingest a source

When the user asks you to ingest or process a source document:

1. Read the source from wiki/raw/<filename> (or from a URL/path the user provides)
2. Discuss key takeaways with the user if needed
3. Write a summary page: wiki/pages/<slug>-source.md
4. Update wiki/index.md — add the new page entry
5. Identify related existing pages and update them (cross-references, new info, contradictions)

• A single source typically touches 5–15 pages
• Note explicitly when new data contradicts an existing claim

1. Append to log: python3 "$SKILL_DIR/scripts/log.py" ingest "<Source Title>" "<brief notes>"

Query the wiki

When the user asks about a topic you've been researching:

1. Search for relevant pages: python3 "$SKILL_DIR/scripts/search.py" "<query>"
2. Read the returned pages (highest-score first)
3. If index.md exists and search returns nothing, read index.md to browse manually
4. Synthesize an answer with citations (page filenames)
5. If the answer is non-trivial and reusable, file it back as a new page in wiki/pages/
6. Append to log: python3 "$SKILL_DIR/scripts/log.py" query "<query summary>" "<key finding>"

Lint the wiki

When the user asks you to health-check or lint the wiki:

1. Run: python3 "$SKILL_DIR/scripts/lint.py"
2. For each issue reported:

• orphan: page has no inbound links — add cross-links from related pages, or mention in index
• missing: listed in index.md but file doesn't exist — remove from index or recreate
• unlisted: file exists but not in index.md — add it to index

1. Read through several pages for contradictions and stale claims
2. Append to log: python3 "$SKILL_DIR/scripts/log.py" lint "Lint pass" "<summary of fixes>"

index.md Format

# Wiki Index

> Last updated: YYYY-MM-DD | Pages: N | Sources: N

## [Category Name]

| Page | Summary | Tags | Updated |
|------|---------|------|---------|
| [Page Title](pages/page-slug.md) | One-line summary | tag1, tag2 | YYYY-MM-DD |

Organize pages by category. Update on every ingest.

log.md Format

## [YYYY-MM-DD] ingest | Source Title
Notes about what was ingested and key pages updated.

## [YYYY-MM-DD] query | Query Summary
Key finding: ...

## [YYYY-MM-DD] lint | Lint pass
Fixed N issues: ...

Scripts

# List all wiki pages with metadata (title, tags, sources, last-updated)
python3 "$SKILL_DIR/scripts/list.py" [--tag TAG] [--sort title|updated]

# Search wiki pages by keyword relevance
python3 "$SKILL_DIR/scripts/search.py" "<query>" [--top N]

# Health-check the wiki (orphans, missing files, unlisted pages)
python3 "$SKILL_DIR/scripts/lint.py"

# Append a log entry
python3 "$SKILL_DIR/scripts/log.py" <ingest|query|lint> "<title>" ["<notes>"]

Environment Variables

• WIKI_DIR — path to the wiki directory (default: ./wiki/). Set this to your preferred persistent location.

Tips

• File answers back: A good analysis you synthesized is a new wiki page — don't let it disappear into chat history
• Contradictions matter: When new info contradicts old, update both pages and note the conflict explicitly
• Prefer wiki over web searches: For topics the user has been researching, the wiki has curated, synthesized knowledge. Use it first.
• Raw sources are immutable: Never modify files in wiki/raw/ — they are the source of truth