wiki

name	wiki
description	Build and maintain a persistent LLM-powered wiki/knowledge base. Use when the user says "wiki", "knowledge base", "ingest this", "add to wiki", "wiki query", "wiki lint", "wiki init", "wiki search", or wants to build a structured, interlinked collection of markdown files from source documents. Based on Karpathy's LLM Wiki pattern.
argument-hint	<command> [args...]
allowed-tools	["Read","Write","Edit","Glob","Grep","Bash","Agent"]

LLM Wiki

Build and maintain a persistent, compounding knowledge base as interlinked markdown files. Instead of re-deriving knowledge on every query (RAG-style), incrementally compile sources into a structured wiki the LLM maintains. The wiki is the artifact; you curate, the LLM does the bookkeeping.

Commands

Parse $ARGUMENTS to determine which operation to run:

Command	Description
init [path]	Scaffold a new wiki at path (default: ./wiki)
ingest <source>	Process a source file/URL and integrate into the wiki
query <question>	Search the wiki and synthesize an answer
lint	Health-check the wiki for issues
search <term>	Full-text search across wiki pages
status	Show wiki stats (page count, source count, recent activity)

If no command is recognized, treat the entire argument as a query.

---

Architecture

The wiki has three layers:

wiki-root/
  raw/            # Immutable source documents (user adds, LLM reads only)
  pages/          # LLM-generated markdown pages (LLM owns entirely)
  index.md        # Content catalog with links and one-line summaries
  log.md          # Append-only chronological record of operations
  schema.md       # Wiki conventions, structure, domain-specific rules

Init

When init is invoked:

1. Create the directory structure above
2. Create schema.md with default conventions (below)
3. Create empty index.md with header and category sections
4. Create empty log.md with header
5. Create raw/ and pages/ directories
6. Print the path and next steps

Default schema.md content:

# Wiki Schema

## Conventions
- All wiki pages live in `pages/`
- Use `[[wikilinks]]` for cross-references between pages
- Each page has YAML frontmatter: title, created, updated, tags, sources
- Page filenames are kebab-case: `machine-learning-basics.md`
- One concept/entity per page — split if a page covers multiple topics

## Page Types
- **source**: Summary of a raw source document
- **entity**: A person, place, organization, or specific thing
- **concept**: An idea, theory, method, or topic
- **synthesis**: Cross-cutting analysis combining multiple sources
- **comparison**: Side-by-side analysis of related items

## Index Format
Categories in index.md: Sources, Entities, Concepts, Syntheses, Comparisons

## Log Format
Each entry: `## [YYYY-MM-DD] operation | Title`

Ingest

When ingest <source> is invoked:

1. Read the source file (or fetch URL content)
2. Create a source summary page in pages/ with frontmatter
3. Identify entities, concepts, and themes mentioned
4. For each entity/concept:
   • If a page exists: update it with new information, note the new source
   • If no page exists: create one
5. Add [[wikilinks]] cross-references in all touched pages
6. Update index.md with new/modified pages
7. Append an entry to log.md
8. Copy/save the raw source to raw/ (if it's a file)
9. Print a summary: pages created, pages updated, key takeaways

Guidelines for ingest:

• Preserve nuance — don't flatten complex arguments into bullet points
• Flag contradictions with existing wiki content explicitly
• Link liberally — every mention of a known entity/concept should be a wikilink
• Include citations back to the source: (Source: [[source-page]])

Query

When query <question> is invoked:

1. Read index.md to identify relevant pages
2. Optionally grep for additional keyword matches across pages/
3. Read the relevant pages
4. Synthesize an answer with citations to wiki pages
5. If the answer is substantial and reusable, offer to save it as a new synthesis page in the wiki
6. Append a query entry to log.md

Lint

When lint is invoked, check for:

1. Contradictions: Pages that make conflicting claims
2. Stale content: Claims superseded by newer sources
3. Orphan pages: Pages with no inbound links
4. Missing pages: Wikilinks that point to non-existent pages
5. Thin pages: Pages with very little content that could be expanded
6. Missing cross-references: Related pages that should link to each other
7. Index gaps: Pages that exist but aren't listed in index.md

Print a report with findings and suggested fixes. Offer to auto-fix what's safe (updating index, adding cross-references, creating stub pages for missing links).

Search

When search <term> is invoked:

1. Grep across all pages/*.md files for the term
2. Print matching pages with context snippets
3. Sort by relevance (title match > heading match > body match)

Status

When status is invoked:

1. Count pages by type (from frontmatter)
2. Count raw sources
3. Show last 5 log entries
4. Show total wikilink count and orphan count
5. Print a summary

Page Template

Every wiki page follows this structure:

---
title: Page Title
type: source|entity|concept|synthesis|comparison
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags: [tag1, tag2]
sources: [source-filename]
---

# Page Title

Brief overview paragraph.

## Key Points
- ...

## Details
...

## Related
- [[related-page-1]]
- [[related-page-2]]

Tips

• Ingest sources one at a time for better integration quality
• After several ingests, run lint to catch gaps
• Save good query answers back to the wiki as synthesis pages
• The wiki is just markdown files — use git for version history
• Works great with Obsidian for browsing (graph view, backlinks)