Clawiki

A literature review system built as Claude Code skills. Give it a topic — it finds papers, downloads PDFs, and builds a cross-referenced knowledge base you can query, compare, and synthesize.

Inspired by karpathy/llm-wiki.

Quick Start

# 1. Clone and open in Claude Code
## Please note only clone the latest main branch commit only otherwise its very heavy!
git clone --depth 1 --filter=blob:none --single-branch --branch main git@github.com:take2rohit/clawiki.git
cd clawiki
claude

# 2. Discover papers on your topic (creates a branch, populates index)
/lit-init "world models for robotics"

# 3. Download PDFs and build wiki pages for all discovered papers
/ingest all

# 4. Ask a question against the knowledge base
/ask "what are the key differences between RSSM and transformer-based world models?"

# 5. Find related papers and pull them in
/related hafner-2023-dreamerv3

# 6. Compare methods side-by-side
/compare hafner-2023-dreamerv3 micheli-2022-iris

# 7. Identify gaps and missing coverage
/gaps

# 8. Health-check the wiki (fixes broken links, removes singleton tags)
/lint

After /ingest all you have a fully cross-referenced wiki. Every subsequent command enriches it.

The Idea

Most LLM + papers workflows are stateless. Upload a PDF, ask questions, close the chat, gone. This is different.

The LLM incrementally builds and maintains a persistent wiki — structured markdown files between you and the raw PDFs. Ingesting a paper doesn't just summarize it: the LLM reads the full PDF, creates a wiki page, updates topic/method pages, adds bidirectional cross-references, and updates the index. Each paper makes the whole knowledge base richer.

Commands

Command	What it does	Files Modified
/lit-init <topic>	Discover papers via web search, create workspace structure, populate index	$BRANCH/index.md (created), $BRANCH/log.md (created), $BRANCH/overview.md (created), bibtex/references.bib (created)
/discover [query]	Web-scan for new papers → adds discovered rows to index (no PDF download)	$BRANCH/index.md, $BRANCH/log.md
/ingest <name|all|discovered>	Download PDF(s), build wiki pages, cross-reference	raw/{name}.pdf (downloaded), $BRANCH/papers/{name}.md (created), $BRANCH/index.md, $BRANCH/log.md, relevant topic/method/benchmark pages
/ask <question>	Query the knowledge base with cited answers	$BRANCH/queries/{slug}.md (created), $BRANCH/log.md
/related <name>	Find related papers → adds discovered rows to index (no PDF download)	$BRANCH/index.md, $BRANCH/log.md
/compare <n1> <n2> ...	Side-by-side method comparison table	$BRANCH/queries/compare-{slug}.md (created), $BRANCH/log.md
/gaps	Find missing coverage, broken cross-references	$BRANCH/log.md (report only, no structural changes)
/lint	Health-check, fix broken links, remove singleton tags	$BRANCH/index.md, $BRANCH/papers/*.md (tag fixes), $BRANCH/log.md
/bibtex [id|all]	Export citations to bibtex/references.bib	bibtex/references.bib, $BRANCH/log.md
/host	Publish the knowledge base to GitHub Pages	Commits, pushes, and builds the current topic branch

$BRANCH = the current git branch name. The wiki directory is always named after the branch (e.g., world_models/ on branch world_models).

Discovery Flow

/discover "JEPA world models 2025"   # scans NeurIPS/ICML/ICLR/arXiv — index only, no downloads
/ingest discovered                   # downloads PDFs + builds wiki pages for all discovered rows

Or use --cite-expand to find papers that directly cite your existing ingested papers:

/discover --cite-expand

Paper Status

discovered  ──>  (ingest downloads PDF + creates wiki page)  ──>  ingested
downloaded  ──>  (ingest reads PDF + creates wiki page)      ──>  ingested

$BRANCH/index.md is the single source of truth for all papers and their status.

Structure

├── raw/                      # Downloaded PDFs (immutable)
├── $BRANCH/                  # Wiki directory (named after the git branch)
│   ├── index.md              # Master index
│   ├── overview.md           # Narrative synthesis
│   ├── log.md                # Activity log
│   ├── papers/               # One page per ingested paper
│   ├── topics/               # Concept pages
│   ├── methods/              # Method descriptions
│   ├── benchmarks/           # Leaderboard tables
│   └── queries/              # Saved comparisons and reviews
└── bibtex/
    └── references.bib

Prerequisites

• Claude Code — CLI, desktop app, or IDE extension
• Nothing else. No API keys, no dependencies.

Clone, open in Claude Code, run /lit-init.

Hosting on GitHub Pages

/host   # commits, pushes, and rebuilds GitHub Pages from the current topic branch

The URL matches the branch name: branch world_models is served at /clawiki/world_models/. Running /host auto-configures Pages. Use /lit-switch + /host to switch which review is live.

Branch policy: main contains only skills and docs. Wiki content and PDFs live on topic branches. See CLAUDE.md for the full rules.

Tips

• $BRANCH/index.md is your home page. Every paper links to [PDF] and [Notes].
• /ingest all after any operation that adds papers (/lit-init, /related, /discover).
• Papers are identified flexibly: filename (hafner-2023-jmlr), P-ID (P001), or arXiv ID (2301.04104).
• raw/ is immutable. Notes live in the branch wiki directory. PDFs are never modified.
• Git-friendly. Everything is plaintext markdown. Commit after each session.
• $BRANCH/log.md is your audit trail — every operation is logged.