Figure-first biomedical paper learning for SKILL.md-compatible agents.
Turn a DOI, PDF, or paper link into a structured learning session: metadata, figures, panel-by-panel explanation, experimental design extraction, critical appraisal, and project-specific takeaways.
Use it with Codex, Claude Code, or another code agent that can load Markdown skills, custom instructions, or project-level agent rules.
The portable core is the root SKILL.md; no Codex-only API is required.
- Codex: add this repository to Codex's supported skills/instructions location.
- Claude Code: add this repository or the root
SKILL.mdto Claude Code's supported project instructions / skill workflow. - Other code agents: point the agent at this repository, or copy
SKILL.mdinto its supported custom-instructions or skill location.
Then ask your agent:
帮我逐图拆解这篇文章，先从背景开始。每次讲完一整张大图，每个 panel 配图和文字。
Most AI paper summaries are too smooth: they tell you what the paper says, but not whether the evidence is strong, whether the controls are enough, or whether the experiment can be adapted to your own project.
Paper Dissect is built around a different reading habit:
Get the figures first. Teach one complete large figure at a time. Extract the experimental design. Then decide what the paper proves, suggests, or overclaims.
See docs/demo-output.md for a full mock example.
**A 图：动物模型设计**
作者先建立肿瘤异种移植或类器官相关模型，再把样本分成 vehicle、
treatment 和 rescue 几组。关键不是“有肿瘤模型”，而是模型来源、
给药路线、时间线和终点指标是否真的支撑作者后面关于机制和转化价值的 claim。
这一图说明：作者在尝试建立肿瘤体内证据链，但还不能把单一异种移植模型
直接等同于真实患者的肿瘤微环境。- Establishes paper identity before interpretation.
- Retrieves figures before deep reading.
- Teaches one complete large figure at a time.
- Uses a fixed teaching format:
好，进入 Figure X-> evidence-chain transition -> figure-level claim ->A 图+ image + explanation -> all panels -> figure summary -> project takeaway. - Explains why each figure appears at that point in the paper: what the previous figure proved, what gap remains, and why the authors changed assay, model, species, or mechanism.
- Crops panels when precision matters.
- Extracts in vitro, ex vivo, and in vivo experimental design.
- Separates proven claims from suggested or overclaimed claims.
- Translates useful methods into the user's research context.
flowchart LR
A["Paper link / DOI / PDF"] --> B["Metadata"]
B --> C["Figure retrieval"]
C --> D["Panel cropping"]
D --> E["Panel-by-panel teaching"]
E --> F["Experimental design extraction"]
F --> G["Critical appraisal"]
G --> H["Project-specific takeaways"]
- Learn a Cell, Nature, Science, NEJM, or PMC-style biomedical paper by figures.
- Prepare for journal club or lab meeting.
- Extract experimental designs from Results sections.
- Compare a paper's evidence chain with your own research question.
- Turn a paper into reusable Chinese learning notes.
- Critique translational medicine papers without inflating weak animal models.
See docs/prompt-recipes.md for ready-to-use prompts.
Common prompts:
帮我逐图拆解这篇文章，先从背景开始。
每次讲完一整张大图，每个 panel 配图和文字。
按 Figure 6 那种格式讲：先说这张图要证明什么，再 A 图、B 图逐 panel 配图解释，最后总结能借鉴到我的项目的实验。
重点学习这篇文章的体内外实验，看哪些能加到我的项目。
客观评价这篇文章的机制和转化证据，和我的项目相比打分。
- Confirm paper metadata: title, DOI, PMID/PMCID, journal, dates.
- Retrieve figures from PMC, publisher static URLs, PDFs, or user-provided files.
- Explain background before Results.
- For Results, discuss one complete large figure per turn.
- Before panel A, explain the evidence-chain transition: previous claim, remaining gap, and why this figure uses the next model or assay.
- Explain every panel as
A 图：标题with the cropped image, what was tested, what changed, what it means, and why the panel matters. - End each large figure with a figure-level logic summary and project-specific experimental takeaways.
- Crop panels precisely when the user wants panel-by-panel study.
- Extract model, groups, controls, time course, readouts, and conclusions.
- Separate what the paper proves from what it only suggests.
- Translate useful designs into the user's research question or project.
This repository is organized as a portable SKILL.md package for Codex, Claude
Code, and other code agents. Any compatible agent can read the root SKILL.md
and follow the workflow.
Supported usage patterns use the same core file:
- Codex: add this repository to Codex's supported skills/instructions location.
- Claude Code: add this repository or the root
SKILL.mdto Claude Code's supported project instructions / skill workflow. - Other code agents: point the agent at this repository, or copy
SKILL.mdinto its supported custom-instructions or skill location.
Codex local-directory example:
mkdir -p ~/.codex/skills
git clone https://github.com/KangjieShen/paper-dissect.git
cp -R paper-dissect ~/.codex/skills/For runtimes with a reload step, restart or refresh the agent so the new skill or instruction file is available.
- Do not use this skill as a replacement for systematic review or expert review.
- Do not continue interpreting from wrong panel crops after a crop error is found.
- Do not treat a single xenograft or in vitro killing assay as direct clinical efficacy.
- Do not treat static spatial gradients as proven temporal growth direction.
- Do not overclaim mechanism when the evidence is only pharmacologic phenocopy.
- Do not let repeated animal figures inflate scientific novelty.
- Do not use non-open-access paper figures in public README assets without permission.
paper-dissect/
├── SKILL.md
├── agents/
│ └── openai.yaml # optional agent metadata example
├── docs/
│ ├── demo-output.md
│ ├── demo-output-oncology.svg
│ ├── hero-panel-dissect.svg
│ ├── prompt-recipes.md
│ └── workflow.svg
└── README.md
MIT