1. Prepare a host and materials
Use an account-configured agent that supports Agent Skills and Python 3.10+. The installer uses the standard library; it does not install models, statistical packages or database accounts.
Create a separate paper workspace containing synthetic or authorized, de-identified materials. Local script execution does not make the whole AI session offline: the model service may receive conversation content.
2. Download the repository
Choose Code → Download ZIP on GitHub, extract it and open a terminal in the directory, or use Git:
git clone https://github.com/duyunzhoudyz/dyz-med-paper-skills.git
cd dyz-med-paper-skills
On Windows, check python --version; substitute py -3 if that is your Python launcher. On macOS/Linux, python3 is often appropriate.
3. Install for your host
Claude Code
python scripts/install_codex.py --target "~/.claude/skills"
python scripts/install_codex.py --target "~/.claude/skills" --check
Codex
python scripts/install_codex.py
python scripts/install_codex.py --check
The default destination is ~/.agents/skills/. The filename remains install_codex.py; --target supports other hosts' documented skill directories. Success reports PASS, the skill count and destination. Different existing content stops installation instead of silently replacing edits.
Install for one paper project
Replace the example with an existing project directory. For Claude Code:
python scripts/install_codex.py --target "/path/to/my-paper/.claude/skills"
For Codex:
python scripts/install_codex.py --project "/path/to/my-paper"
Keep the same target arguments and add --check to verify. Choose personal or project scope for each host to avoid duplicate entries.
4. Start a task
Open the paper workspace in your host and reload its skills. Describe the task naturally:
Use dyz-med-orchestrator to inspect the available manuscript materials, identify the most useful next step and help with that task. List missing inputs separately.
Explicit invocation is /dyz-med-orchestrator in Claude Code or $dyz-med-orchestrator in Codex. File checks do not certify that a host has loaded or routed a skill correctly; verify that in the host.
Upgrade and troubleshoot
Run --check with the original destination arguments first. To upgrade from a new checkout, provide a new backup directory outside both the source and installation directories:
python scripts/install_codex.py --upgrade --backup "/path/to/new-backup"
For Claude Code, also keep --target "~/.claude/skills". Manual installation must include complete skill directories and the repository's LICENSE, NOTICE and COMMERCIAL-LICENSE.md, not just SKILL.md files.