Start here

▁▂▃▅▃▂▁

Getting started

Install an AI agent — Claude Code or Codex — in the terminal, add the Open Science Skills, and point it at a project. A one-time setup, with no prior command-line experience assumed.

Overview

The aim is a setup where an AI agent reads the files in a folder and runs reusable checks over them, instead of a chat window you paste into. You install an agent, add the Open Science Skills, and point it at a project. After that, checking a reference list, or testing whether a source supports a claim, is a single command.

These instructions use the terminal on purpose, not just because it's the simplest way in. A terminal agent is scriptable — a skill is just a command, so you can chain them, automate them, or run them unattended. It works the same way over SSH on a remote server or an HPC cluster as it does on your laptop. And it isn't locked to one vendor's app: every demo on this site assumes the terminal, and the steps below work whether the agent running it is Claude Code or Codex.

You can also run Claude Code in an editor like VS Code or JetBrains, or through the web and desktop apps, if you'd rather not use the terminal at all. The steps below cover Claude Code first, since it's what the rest of this site uses, then Codex as an alternative.

1

Install Claude Code

Claude Code is the AI tool that runs the skills. Install it with one command in your terminal:

# macOS, Linux, or WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows (PowerShell)
irm https://claude.ai/install.ps1 | iex

Other install options are in the docs at code.claude.com/docs. Check it worked with claude --version.

2

Sign in

Open a terminal and run:

claude

The first time, it walks you through signing in with your Claude account in the browser. After that it stays logged in.

3

Add the Open Science Skills

The reference and source-claim checks come from the Open Science Skills plugin. Register the marketplace and install it:

claude plugin marketplace add scdenney/open-science-skills
claude plugin install oss@open-science-skills

Restart Claude Code (or run /reload-plugins) so the skills load. To confirm, type /help and look for the oss: commands, such as /oss:citation-check.

4

Open a project and call a skill

Move into the folder you want to work in and start Claude Code there:

cd path/to/your/project
claude

Now you can run a skill two ways. By its slash command:

/oss:citation-check manuscript.md references.bib

Or just by asking in plain language, for example "check the references in manuscript.md against references.bib." Claude Code reads the files in the folder, so point it at real files.

5

Using Codex instead alternative

The Open Science Skills ship a parallel, Codex-native library of 40 of the toolkit’s 43 skills. Codex isn't a drop-in replacement for Claude Code — there's no plugin marketplace, and invocation looks different — but the skills themselves cover the same ground.

Install the Codex CLI itself first; that step isn't part of this toolkit, so follow OpenAI's own instructions at developers.openai.com/codex/skills. Once Codex is installed and signed in, clone the Open Science Skills repo and symlink its skills into Codex's user-wide skills folder:

git clone https://github.com/scdenney/open-science-skills
cd open-science-skills
mkdir -p "$HOME/.agents/skills"
for skill in "$PWD"/codex/*/; do
  ln -sfn "${skill%/}" "$HOME/.agents/skills/$(basename "$skill")"
done

Where Claude Code uses a slash command like /oss:citation-check, Codex invokes the same skill with a $ prefix:

$citation-check manuscript.md references.bib

Codex can also pick up a skill implicitly from how you describe the task, without typing the $ form. The two libraries track each other closely but aren't identical — presubmit has no Codex equivalent, and fable-orchestrate is named 46-orchestrate on the Codex side. See codex/README.md for the full catalog, including repo-scoped and single-skill installs.

6

Build a knowledge base optional

This step is only for the source-claim check, which reads your sources to see whether they support a claim. The reference check does not need it.

Keep each source's PDF in sources/og/ and a Markdown version in sources/md/, which is what the checker reads. The research-repo skill sets this folder structure up for you, and process-source converts a PDF to Markdown.

/oss:research-repo            # scaffold the sources/ folders
/oss:process-source paper.pdf # convert one PDF into the knowledge base