llm-wiki

name	llm-wiki
description	Maintain a personal team knowledge base using the LLM Wiki pattern — incremental ingest, query, and lint operations on a layered wiki architecture
category	knowledge-management
tags	["wiki","knowledge-base","documentation","research","note-taking","obsidian"]
version	1.0.0

LLM Wiki

A pattern for building and maintaining a structured, interlinked collection of markdown files — a personal or team knowledge base — using LLM agents. The wiki is maintained incrementally alongside code work.

When to Use

Invoke this skill when you want to:

• Capture research, decisions, or references from web searches or documents into a persistent knowledge base
• Query accumulated knowledge to answer questions that require synthesizing multiple sources
• Maintain a log of what the wiki has grown to contain
• Onboard a new agent or human to a topic by pointing them at the wiki

The wiki is opt-in. Projects that don't need it are unaffected. Projects that do can build it up over time.

Setup

When activating this skill on a fresh project, create the wiki directory structure and seed files:

mkdir -p wiki raw
touch wiki/index.md wiki/log.md

wiki/index.md Template

# Wiki Index

> Content catalog for the LLM Wiki. Maintained by agents after each ingest operation.

## Topics

<!-- Add entries as you create pages:

- [[topic-name]] — Brief description

-->

## Recent Additions

<!-- Chronological list of new pages:

- [[YYYY-MM-DD]] [[page-name]] — Why it was added

-->

wiki/log.md Template

# Wiki Log

> Append-only chronological record of wiki operations. Never edit past entries.

## Usage

Each entry follows this format:

YYYY-MM-DD

ingest

• Sources:
• Pages created:
• Pages updated:
• Summary:

query

• Question:
• Answered by:
• Summary:

lint

• Issues found:
• Issues fixed:
• Summary:

---

## Architecture

The wiki operates in three layers:

┌─────────────────────────────────────────────────────────────┐ │ Raw Sources (immutable, timestamped inputs) │ │ raw/ │ │ PDFs, web captures, transcript files, etc. │ └─────────────────────────────────────────────────────────────┘ │ Ingest ▼ ┌─────────────────────────────────────────────────────────────┐ │ Wiki (LLM-maintained markdown pages) │ │ wiki/ │ │ Topic pages, concept notes, decision records, etc. │ │ Linked via [[wiki-links]] │ └─────────────────────────────────────────────────────────────┘ │ Synthesize ▼ ┌─────────────────────────────────────────────────────────────┐ │ Schema (instructions for AGENTS.md) │ │ Key facts, conventions, decisions that agents must follow │ └─────────────────────────────────────────────────────────────┘

### Raw Sources (`raw/`)

Immutable source material. Never edit after creation — this preserves provenance.

**Naming convention:**

YYYY-MM-DD-.

Example: `2024-01-15-attention-mechanism-paper.pdf`

### Wiki Pages (`wiki/`)

Markdown files maintained by agents. Organized by topic, linked via `[[double-bracket]]` links (standard in Obsidian, supported by most markdown link tools).

**Naming convention:**

-.md

Example: `attention-mechanism.md`, `transformers-vanishing-gradients.md`

**Frontmatter (required):**
```yaml
---
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags: [tag1, tag2]
---

# Topic Title

...

Schema Layer

Facts, conventions, and decisions extracted from the wiki and incorporated into AGENTS.md or project documentation. This is the loop that closes: wiki → actionable agent knowledge.

---

Operations

Ingest

Add new sources and update the wiki with their content.

When to ingest:

• After a research session (web search, paper reading, documentation review)
• When adding a new topic page
• When updating an existing page with new information

Steps:

1. Create raw source entry (if source is new):

   cp <source-file> raw/$(date +%Y-%m-%d)-<slug>.<ext>
   

2. Read and extract key information from the source.

3. Create or update wiki page at wiki/<topic>.md:

   • Use the frontmatter template
   • Link to related pages with [[wiki-links]]
   • Include a "Sources" section with references

4. Update wiki/index.md:

   • Add the new topic entry
   • Add a "Recent Additions" entry with date

5. Append to wiki/log.md:

   ### ingest
   - **Sources:** <!-- -->
   - **Pages created:** <!-- -->
   - **Pages updated:** <!-- -->
   - **Summary:** <!-- -->
   

Query

Answer a question by consulting the wiki.

When to query:

• Before starting work on a topic you haven't touched recently
• When you need to recall a decision made in a previous session
• When onboarding and you need to understand the project context

Steps:

1. Read wiki/index.md to find relevant topic pages.

2. Read the relevant wiki pages for the topic.

3. Synthesize an answer from the wiki content.

4. Append to wiki/log.md:

   ### query
   - **Question:** <!-- -->
   - **Answered by:** <!-- pages consulted -->
   - **Summary:** <!-- answer or "no answer found" -->
   

Lint

Health-check the wiki for issues.

When to lint:

• Before a release or handoff
• When the wiki starts to feel disorganized
• As a periodic maintenance step

Steps:

1. Check for orphaned pages — files that nothing links to:

   # Find markdown files with no [[links]] pointing to them
   find wiki -name '*.md' -exec grep -L "^\\[\\[" {} \;
   

2. Check for broken links — [[wiki-links]] that point to non-existent pages:

   grep -roh '\[\[[^]]\+\]\]' wiki/ | sed 's/\[\[\(.*\)\]\]/\1/' | while read page; do
     [ -f "wiki/$page.md" ] || echo "MISSING: $page"
   done
   

3. Check frontmatter — all pages should have created, updated, and tags.

4. Fix issues found during the check.

5. Append to wiki/log.md:

   ### lint
   - **Issues found:** <!-- count -->
   - **Issues fixed:** <!-- count -->
   - **Summary:** <!-- overall health -->
   

---

Page Templates

Concept Page

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags: [concept, area]
---

# Concept Name

One-line definition or summary.

## Background

Why this concept matters.

## Key Points

- Point 1
- Point 2

## Related

- [[related-page-1]]
- [[related-page-2]]

## Sources

Decision Record

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags: [decision, area]
---

# Decision: Short Title

**Status:** Proposed | Accepted | Deprecated
**Date:** YYYY-MM-DD

## Context

What situation forces this decision?

## Decision

What was decided?

## Consequences

What changes because of this decision?

## Related

- [[related-page]]

---

Optional Tooling

These integrations are optional — the wiki works as plain markdown files.

Obsidian

The wiki format is fully compatible with Obsidian. Just open the wiki/ folder as a vault.

Recommended plugins:

• obsidian-markdown-links — render [[wiki-links]] as navigable links
• obsidian-dataview — query pages by tags, created date, etc.

qmd (Quick Markdown Search)

A lightweight search engine for markdown files. Index your wiki:

# Install qmd (if available)
pip install qmd

# Index the wiki
qmd index wiki/

# Query
qmd search "attention mechanism"

Marp Slides

Generate presentations from wiki content:

# Install Marp CLI
npm install -g @marp-team/marp-cli

# Convert a page to slides
marp wiki/attention-mechanism.md --output slides/

---

Integration with AGENTS.md

When the wiki is established, add conventions to AGENTS.md:

## Wiki

This project maintains a knowledge base at `wiki/`. Before starting work on a topic, query the wiki:

```bash
# Check for relevant pages
grep -r "topic" wiki/ --include="*.md"

After research or significant decisions, ingest into the wiki per the llm-wiki skill.

---

## References

- [Karpathy's LLM Wiki (original gist)](https://github.com/karpathy/llm-wiki) — Source of this pattern
- [Obsidian](https://obsidian.md/) — Recommended local interface
- [Marp](https://marp.app/) — Markdown presentation engine