arxiv-paper-html-cn-notes-skill

This repository stores reusable AI-agent skills. The current skill is:

• arxiv-paper-html-notes: generates polished Chinese static HTML notes from one or more arXiv papers.

What It Does

arxiv-paper-html-notes helps an AI agent produce:

• single-paper deep reading pages
• multi-paper review pages
• Chinese literature-review webpages
• meeting or group-presentation style paper analysis pages

The skill bundles:

• SKILL.md: workflow and output rules
• references/: content schema, visual style, and verification checklist
• scripts/: arXiv metadata, PDF text extraction, PDF figure rendering, and HTML assembly helpers
• assets/html-template/: static HTML/CSS template

Example Output

Example generated page:

• APOLLO: A Virtual Patient Foundation Model

Tool Compatibility

Tool	Directly usable from this repo?	Status	Notes
Codex	No, not from the repo root automatically	Supported after install	Copy arxiv-paper-html-notes/ to ~/.codex/skills/.
Claude Code	Yes, project-level version included	Supported	This repo includes .claude/skills/arxiv-paper-html-notes/.
GitHub Copilot	Not as a SKILL.md skill	Supported through instructions	This repo includes .github/copilot-instructions.md and .github/instructions/arxiv-paper-html-notes.instructions.md.
Generic agents	Partially	Supported through AGENTS.md	Agents that read AGENTS.md can find the canonical skill.

Versions in This Repo

• arxiv-paper-html-notes/: canonical, tool-neutral skill package.
• .claude/skills/arxiv-paper-html-notes/: Claude Code project-level skill copy.
• .github/copilot-instructions.md: repository-wide GitHub Copilot entrypoint.
• .github/instructions/arxiv-paper-html-notes.instructions.md: Copilot task-specific instruction file.
• AGENTS.md: generic agent entrypoint.

If you edit the canonical skill, sync the Claude Code copy:

rm -rf .claude/skills/arxiv-paper-html-notes
cp -R arxiv-paper-html-notes .claude/skills/

Install

Codex

Install as a personal Codex skill:

mkdir -p ~/.codex/skills
cp -R arxiv-paper-html-notes ~/.codex/skills/

Then ask Codex for something like:

Use the arxiv-paper-html-notes skill to generate a Chinese HTML note for https://arxiv.org/abs/....

Claude Code

Project-level use is already prepared:

.claude/skills/arxiv-paper-html-notes/

For global use across projects:

mkdir -p ~/.claude/skills
cp -R arxiv-paper-html-notes ~/.claude/skills/

Then ask Claude Code:

Use the arxiv-paper-html-notes skill to create a Chinese review page for these arXiv papers: ...

GitHub Copilot

Copilot does not automatically load SKILL.md as a skill. Use the included repository instructions:

.github/copilot-instructions.md
.github/instructions/arxiv-paper-html-notes.instructions.md

In a Copilot chat, ask:

Use the arxiv-paper-html-notes workflow in this repo to generate a Chinese HTML paper note for ...

Python Dependencies

The helper scripts use Python 3.10+.

Install optional packages when needed:

python -m pip install pypdf pypdfium2

Script usage:

python arxiv-paper-html-notes/scripts/extract_arxiv_metadata.py <arxiv-url-or-id> --out metadata.json
python arxiv-paper-html-notes/scripts/extract_pdf_text.py <pdf-or-url> --out paper.txt
python arxiv-paper-html-notes/scripts/render_pdf_figures.py <pdf> --page 6 --out assets/paper_figure.png
python arxiv-paper-html-notes/scripts/assemble_html.py paper_notes.json --out index.html --asset-dir assets

Usage Tips

• Always verify paper titles, arXiv versions, dataset numbers, and evaluation metrics against the paper or official source.
• Mark code/model availability as unknown unless confirmed from arXiv comments, official project pages, GitHub, or Hugging Face.
• Use generator_name in paper_notes.json to match the calling tool, such as Codex, Claude Code, or GitHub Copilot.
• For one paper, produce a deep single-paper page and omit empty comparison sections.
• For two to six papers, include paper navigation, comparison, synthesis, references, and credits.
• If figure extraction is unreliable, use a self-drawn CSS method diagram and leave figure_path empty.

Prompt Templates

Single-paper deep note:

Use the arxiv-paper-html-notes skill to generate a Chinese static HTML deep reading note for:
https://arxiv.org/abs/....

Output to notes/<paper-slug>/index.html.
Include paper metadata, dataset details, method diagram, validation results, open-source status, and a concise Chinese summary.

Multi-paper review page:

Use the arxiv-paper-html-notes skill to create a Chinese multi-paper review page for these arXiv papers:

1. https://arxiv.org/abs/....
2. https://arxiv.org/abs/....
3. https://arxiv.org/abs/....

Group them by research direction, compare backbone, data scale, downstream tasks, clinical validation, and open-source status.
Output to notes/<topic-slug>/index.html.

Meeting or group-presentation style page:

Use the arxiv-paper-html-notes skill to prepare a Chinese reading-group HTML page for these papers:
<arXiv links>

Focus on motivation, method pipeline, why the validation is convincing or weak, limitations, and discussion questions.

Multi-Paper Tips

• Use two to six papers per page. More than six usually becomes too dense; split by disease, organ, modality, year, or model family.
• Put all arXiv links in one request when you want a comparison table and shared synthesis.
• Tell the agent the comparison angle if you already know it, for example pathology foundation models for lung cancer, virtual patient models, or radiology multimodal agents.
• Ask the agent to normalize paper names into short labels, such as APOLLO, Virchow2, or UNI.
• Ask for a stable output folder, such as notes/pathology-foundation-models/index.html, so assets and links stay organized.
• For review pages, require a final synthesis section covering shared trends, strongest evidence, weak spots, and future directions.
• If papers are from different domains, ask the agent to explain why they are grouped together; otherwise the comparison table can become shallow.
• For fresh papers, ask the agent to re-check official sources before judging code/model availability.

Output Quality Checklist

• paper_notes.json follows references/content_schema.md.
• Every local image referenced by figure_path exists under the output assets/ folder.
• Each paper has at least one arXiv link and any confirmed project/code/model links.
• Dataset numbers, model names, and benchmark metrics are traceable to the paper text or official sources.
• The HTML opens locally and does not have obvious desktop or mobile overflow.
• The footer uses the right generator_name and keeps Paper Analysis & Curation by CocoSun.