Paper Repro Python Skill

A reusable skill for reproducing research papers in Python with a strict source-priority workflow: read local TeX sources first when available, use full-fidelity PDF-to-Markdown extraction only as fallback, then implement paper-specific code and maintain bilingual project documentation.

Compatible with: Codex, Claude Code, OpenClaw

What This Skill Does

• Prioritize local TeX sources (.tex, .bib, styles, figures) for paper understanding and reproduction.
• Fall back to PDF extraction only when TeX sources are missing or incomplete.
• Preserve original paper content without summarization, paraphrasing, or rewriting scientific statements.
• Plan and implement reproduction in Python based on source materials, not memory.
• Enforce modular engineering principles: low coupling, high cohesion, clear boundaries.
• Encourage module splitting to avoid monolithic files; keep one source file under ~200 lines whenever practical.
• Generate and maintain both README.md (English) and README_zh-CN.md (Chinese).
• Require README files to start with paper metadata: title, authors (with affiliations and emails), and abstract.
• Require generated images/figures to be embedded in both README files.
• Prefer one chart per image file; only use multi-panel combined figures when comparison across panels is necessary.

Typical Trigger

Use this skill when your request includes one or more of these goals:

• Reproduce a paper from a local TeX project or a PDF.
• Extract complete paper content to Markdown when PDF fallback is needed.
• Build a cleanly modular Python reproduction project.
• Update project docs with bilingual README files and experiment figures.

Example prompt:

Use $paper-repro-python. If TeX files exist in the folder, read TeX first; otherwise extract this paper PDF to full Markdown (no summarization). Then reproduce in Python with clear modules (avoid monolithic files), and update README.md + README_zh-CN.md with generated figures.

Installation

Codex

Option A: Local install (copy to global Codex skills)

Windows (PowerShell):

Copy-Item -Recurse -Force .\paper-repro-python $env:USERPROFILE\.codex\skills\

macOS/Linux (bash/zsh):

mkdir -p "$HOME/.codex/skills"
cp -R ./paper-repro-python "$HOME/.codex/skills/"

Then restart the Codex client.

Option B: Install from GitHub with $skill-installer

1. Ensure the repo path points to a folder containing SKILL.md.
2. In Codex chat, send one of the following commands.

If the skill is at repo root:

Use $skill-installer and install from:
https://github.com/celynnmoonlight/paper-repro-python/tree/main

If the skill is in a subfolder:

Use $skill-installer and install from:
https://github.com/celynnmoonlight/paper-repro-python/tree/main/skills/paper-repro-python

If you prefer repo/path form:

Use $skill-installer and install from repo openai/skills path skills/.curated/<skill-name>

3. After installation, restart Codex to pick up new skills.

Claude Code

Copy the skill folder to your project's .claude/skills/ directory:

mkdir -p .claude/skills
cp -R paper-repro-python .claude/skills/

Or install globally:

mkdir -p "$HOME/.claude/skills"
cp -R paper-repro-python "$HOME/.claude/skills/"

OpenClaw

Copy the skill folder to OpenClaw's skills directory:

cp -R paper-repro-python ~/.openclaw/skills/

Or use OpenClaw's skill installer if available.

Directory Layout

paper-repro-python/
  SKILL.md
  README.md
  README_zh-CN.md
  .gitignore
  LICENSE
  agents/
    openai.yaml

Notes

• Keep SKILL.md as the source of behavior.
• Keep agents/openai.yaml aligned with SKILL.md metadata.
• If TeX and PDF disagree, document the discrepancy and prefer the source that is more complete for the targeted claims.
• If extraction quality is limited by scanned PDFs/OCR, mark uncertain text explicitly.