An agent skill for building a source-grounded wiki from a curated collection of research papers. It scaffolds the project, helps collect open-access sources, guides paper-by-paper analysis and cross-paper synthesis, checks citations and evidence locators, and builds an offline static site with search and KaTeX math.
The skill lives in literature-wiki/. Its SKILL.md is the agent entry point; references/ holds writing guidance, scripts/ holds the wiki tools, and templates/ holds the editing contract and README copied into new projects. Five small command skills sit next to it, /create-wiki, /add-papers, /wiki-exams, /serve-wiki and /update-wiki, so the common tasks are one command away. It works with Codex and Claude skill directories.
These screenshots show Learning Without Backprop, a literature wiki built with this workflow. Click an image to see it at full size.
Home page · Start with the textbook or browse the paper collection.
Textbook · A guided introduction to the topic, grounded in source papers.
Paper page · Source details and navigation into the paper analysis.
Paper figures · Original diagrams appear inline with their source and caption.
Figure 1 in the last screenshot is from Frenkel, Lefebvre and Bol (2021), © 2021 the authors, published under CC BY 4.0. The screenshot shows it within the example wiki.
Clone the repository, install Python 3.10 or newer, then install the script dependencies:
git clone https://github.com/Fuminides/literature-wiki-skill.git
cd literature-wiki-skill
python3 -m pip install -r requirements.txtThe source and figure tools also use pdftotext, pdfinfo, and pdftoppm from Poppler. detex and Ghostscript (gs) are optional. Node.js is needed only for the optional math check.
Copy the skill and its commands into the agent's skill directory. For Claude Code:
mkdir -p ~/.claude/skills
cp -R literature-wiki create-wiki add-papers wiki-exams serve-wiki update-wiki ~/.claude/skills/For Codex, use ~/.codex/skills/ in both commands. The command folders must sit next to literature-wiki, whose scripts and guides they use. You can also use the scripts directly from the cloned repository without installing the agent skill.
In Claude Code, type /create-wiki followed by whatever you have: a paper list, a file, a folder of PDFs, or just a topic.
/create-wiki causal discovery from observational data, papers in ~/reading/causal.bib
/create-wiki a wiki on hypernetworks from the PDFs in ~/papers/hypernetworks/
/create-wiki program synthesis with LLMs
It works out the wiki's name, topic and directory and builds the corpus from your input. Given only a topic, it proposes a reading list and waits for your approval. It confirms everything once, then runs every phase: downloads, paper pages, synthesis, textbook, site and exams. A full build is long, so it reports after each phase and pilots the first pages before writing the rest.
| Command | What it does |
|---|---|
/create-wiki [topic, papers, file or folder, directory] |
Creates a wiki from whatever you have, as above. |
/add-papers [papers] [wiki] |
Adds papers to an existing wiki and carries them through the paper pages, synthesis, textbook, exams and site. |
/wiki-exams [level] [count] [wiki] |
Writes or extends the easy, medium and hard exams, piloting a few questions first. |
/serve-wiki [wiki] [--port N] |
Starts the site with the Ask Claude panel and graded exams, rebuilding it if the wiki changed, and gives you the URL. |
/update-wiki [wiki] |
Brings a wiki made with an older version of the skill up to date (chat panel, exams, fixes, README) and keeps its local changes. |
The commands run only when you type them. Plain requests work too, because the literature-wiki skill picks up requests such as “create a wiki about causal discovery from these papers in ./causal-wiki”. In Codex, ask for a command's skill by name, for example “use create-wiki with ~/reading/causal.bib”.
To scaffold a project directly:
python3 literature-wiki/scripts/init_wiki.py ./causal-wiki \
--name "Causal Discovery" \
--topic "Methods and evidence for causal discovery from observational data."
cd causal-wiki
python3 .tools/lint_wiki.py
python3 .tools/build_site.pyThe scaffold creates papers.csv for your curated corpus, wiki/ for the pages, .tools/ for the scripts and site assets, both AGENTS.md and CLAUDE.md with the project's evidence and editing rules, and a README.md that tells readers how to use the site, the chat and the exams. Fill papers.csv before the source collection phase. The generated site is in site/ and can be opened directly from disk.
To ask questions while you read, run /serve-wiki, or python3 .tools/serve_site.py, and open the URL it prints (http://127.0.0.1:8765 unless that port is taken). Every page then has an "Ask Claude" panel that knows which page you are on and any passage you highlight, and answers with links into the wiki. It needs Claude Code installed and logged in.
The site can also hold easy, medium and hard exams that test understanding of the field rather than recall of individual papers. Multiple-choice questions are graded in the browser. When the site is served this way, Claude grades written answers and explains your mistakes. The server also counts how often each question has been answered, and once you have practised a level enough, Claude writes new questions for it. If the wiki has nothing significant left to ask at that level, the exam page says so instead.
The scripts contact arXiv or Semantic Scholar only when you run the source discovery and download commands. Download only copies you are allowed to use. The vendored KaTeX files retain their own MIT license.
The skill and original tools are available under the MIT License.