Skip to content

Medical writing Agent Skills pack ships 32 skills for Claude Code

A public beta repository packages medical paper workflows for Chinese clinicians as installable Agent Skills for Claude Code and Codex.

Draftship Editorial3 min read

A 32-skill pack for medical papers

A public GitHub repository named duyunzhoudyz/dyz-med-paper-skills packages 32 medical writing Agent Skills for Chinese doctors and medical graduate students. The project was created on 2026-11-01 and had 370 stars and 3 forks as of 2026-11-14, 13 days later. It is written mainly in Python and lists itself as covering PubMed literature search, statistical analysis, Chinese and English writing, journal selection and submission, and answering peer review. The skills are built for Claude Code, Codex and other agents that support Agent Skills. The project calls itself a public beta.

What the 32 skills cover

The skills follow a paper through a working sequence rather than a single prompt. According to the project's own README, they can help narrow a research question from existing case data, design Chinese and English search strategies, interpret statistical output and separate statistical difference from clinical meaning, draft sections, work through supervisor feedback, draft point-by-point reviewer replies, and shortlist journals by topic, article type, budget and requirements.

It lists observational studies, randomized trials and systematic reviews among its writing scenarios, and mentions PRISMA for systematic reviews. There is also an orchestrator skill, dyz-med-orchestrator, that a user can call to have the agent look at a folder of materials and say what should be tackled first.

The project is explicit that this is not one-shot paper generation. It says the assistant chooses a next step from available material, and that the paper is not completed automatically in a single call.

How installation and invocation work

Each skill is a directory containing a SKILL.md file. A Python installer script, scripts/install_codex.py, copies all 32 skills plus licence files to a skills directory, then prints PASS, the skill count and the install path. It is offline, uses only the standard library, does not install models or R and Python packages, and does not change host settings or other skills. If a skill of the same name already exists with different content, it stops.

Personal-level paths are ~/.claude/skills/ for Claude Code and ~/.agents/skills/ for Codex. Project-level paths are .claude/skills/ and .agents/skills/, set with --target for Claude Code or --project for Codex. Python 3.10 or later is required, and a --check flag verifies that installed files match. Explicit calls look like /dyz-med-orchestrator in Claude Code and $dyz-med-orchestrator in Codex.

Limits, data handling and licensing

The status section is unusually candid for a young repository. It says continuous writing quality in real use, actual behaviour across different agents, and independent evaluation of the medical content are not fully done. Writing results, it notes, depend on the model, the material and how the task is described.

The built-in online search tool currently covers PubMed only. Searching Chinese sources such as CNKI, Wanfang and Weipu can be planned, but access needs the relevant permissions and the installer does not bring database accounts. Statistics need appropriate software and a confirmed analysis plan. Word export is described as suitable for a basic review draft, with complex formatting and tracked changes needing other tools.

On privacy, the source warns that running local scripts does not make the whole AI session offline. Model services may receive content supplied in the session, so identifiable patient information, hospital credentials and unapproved material should stay out of it.

The licence is PolyForm Noncommercial 1.0.0, and uses beyond its scope need separate written commercial permission. The examples in the repository are described as synthetic teaching material, not user feedback or published papers.

What this suggests

This is likely most useful as a packaging pattern rather than as a product. A specialist writes down a workflow as a set of named skill folders and a small installer, and a coding agent picks it up in the host it already runs in. That approach probably travels to other narrow, document-heavy domains where the expertise, not the code, is the hard part. It also suggests the review burden does not move: the project itself still puts final research judgement and the submitted manuscript in human hands.

Building apps with AI

For people building web apps through Draftship, nothing here changes what the platform supports, and Draftship is not affiliated with the project. The transferable lesson is about prompt and instruction design: a narrow, well-scoped task description gets better output than a broad request. Draftship's guide to writing effective prompts covers that ground for app builds.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.