Llm wiki

Skill kyuna0312/context_forge/skills/llm-wiki

This skill should be used when the user says "build a wiki", "maintain a wiki", "ingest docs into wiki", "query my wiki", "set up llm wiki", "wiki-based knowledge base", "stop re-reading docs every session", "persistent knowledge base", or "compress my docs into wiki pages". Do NOT use for source code (grep the code — a wiki copy of it rots) or for one-off questions about a document.From its SKILL.md

Install

npx -y skills add kyuna0312/context_forge --skill llm-wiki

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.3 KB, as published. Nobody here has run it

LLM Wiki

Build and maintain a persistent wiki Claude can reference instead of re-ingesting raw documents each session. Derived from Karpathy's LLM Wiki pattern.

Why it works: Raw docs reloaded every session = wasted tokens. A wiki is a distilled, pre-summarized artifact — same knowledge, 10x lower token cost.

Architecture

raw-sources/   (immutable originals)
  └── docs, PDFs, changelogs, README files

wiki/          (LLM-maintained markdown pages)
  ├── index.md        ← content catalog
  ├── log.md          ← append-only ingest/query log
  └── [topic].md      ← one page per concept

CLAUDE.md      (tells Claude: "reference wiki/ instead of raw-sources/")

Setup

Create wiki directory at project root or ~/.claude/wiki/ for global:

mkdir -p wiki
printf '# Wiki Index\n\n## Pages\n' > wiki/index.md
printf '# Ingest Log\n' > wiki/log.md

Add to CLAUDE.md:

## Knowledge Base
Wiki at `wiki/`. Reference wiki pages instead of raw source files.
Do not re-read raw docs unless wiki page is missing or stale.

Operations

Ingest: Add New Source to Wiki

When user provides a doc, URL, or file to ingest:

Read the source
Identify 3-5 key concepts worth persisting — if the source is smaller than the wiki page would be (a short README, a config file), skip it: re-reading it is cheaper than maintaining a copy
Create or update wiki/[topic].md per concept (one page per topic)
Update wiki/index.md with new pages
Append to wiki/log.md: [DATE] INGEST: [source] → [pages updated]

Per wiki page format:

# [Topic]
_Last updated: [DATE] | Source: [origin]_

## Summary
[2-3 sentence distillation]

## Key Points
- [point 1]
- [point 2]

## Related
- [[other-wiki-page]]

Query: Answer from Wiki

When user asks a question:

Search wiki/index.md for relevant pages
Read matching wiki pages
Synthesize answer with citations: (wiki/[page].md)
Create a new page only if answering required going back to raw sources — that gap is the signal a page is missing. Synthesis alone is not a page.
Log to wiki/log.md only when the wiki changed ([DATE] INGEST/UPDATE: …). Don't log read-only queries — a write per question is ceremony, and the log would outgrow the wiki.

Lint: Health Check

Periodically check wiki integrity:

Contradictions:  pages with conflicting claims
Stale pages:     source file's mtime is newer than the page's "Last updated"
Orphaned pages:  not referenced in index.md
Missing links:   [[references]] with no matching file

A stale page is worse than no page — it answers confidently and wrong. Re-ingest or delete; never leave it.

Report format:

WIKI HEALTH
===========
Pages: [N]
Last ingest: [DATE]
Issues found:
- [issue 1]
- [issue 2]
Recommendation: [action]

Token Savings Estimate

Approach	Tokens per session
Re-read raw 50-page doc	~15,000 tokens
Reference 3 wiki pages	~600 tokens
Savings	~96%

When NOT to Build a Wiki

Source code — the code is its own source of truth; a wiki copy drifts the moment the code changes. Wiki decisions and concepts, grep the code.
Sources smaller than their page (short README, config file) — re-reading the original is cheaper than maintaining a copy.
One-off questions — answer from the doc directly; a wiki pays off on the second read, not the first.
Anything with a maintained upstream reference (official API docs you can fetch) — wiki only the project-specific deltas and gotchas.

Integration with context_forge

Run /context_forge:estimate-tokens after wiki setup to measure baseline
Add the wiki/ reference to CLAUDE.md via /context_forge:optimize-claudemd

Additional Resources

references/wiki-patterns.md — Page templates and multi-project wiki setups