Follow this workflow end-to-end unless the user explicitly asks to skip steps
name: paper-repro-python
by celynnmoonlight · published 2026-03-22
$ claw add gh:celynnmoonlight/celynnmoonlight-paper-repro-python---
name: paper-repro-python
description: This skill should be used when the user asks to "reproduce a paper", "implement paper methods in Python", "extract paper content to Markdown", or works on paper reproduction tasks. Use for TeX-first extraction, modular Python implementation, and bilingual documentation.
version: 1.0.0
metadata:
openclaw:
emoji: "📄"
---
# Follow this workflow end-to-end unless the user explicitly asks to skip steps
1) Intake and scope
- If usable TeX source files are present, use TeX as the primary source for reproduction.
- If TeX is absent or incomplete for key content, fall back to PDF extraction only for missing parts.
2) Source extraction (TeX-first, PDF fallback)
- Parse and read the main TeX project structure first (`main.tex` or equivalent entry file and includes).
- Preserve original scientific wording when converting relevant content to Markdown notes.
- Resolve equations, theorem blocks, citations, and appendices from source files whenever possible.
- Record unresolved include/bibliography issues explicitly; do not invent missing content.
- Extract paper content page by page into Markdown, preserving the original wording.
- Do not summarize, paraphrase, or rewrite scientific statements.
- Preserve structure faithfully:
- Title, authors, affiliations, abstract, sections, subsections.
- Equations (LaTeX-friendly when possible), theorem/lemma/proposition blocks.
- Tables, figure captions, references, appendices, footnotes.
- If a PDF is scanned or partially unreadable:
- Run OCR and mark uncertain spans clearly.
- Never silently invent missing text.
- Include image references/placeholders when figures cannot be represented as plain text.
- Produce one primary output file such as `paper_fulltext.md`.
3) Extraction quality checks
- Section/headings coverage matches the TeX project or PDF source used.
- Key equations and algorithm blocks are present.
- References and appendices are included if present in the source.
4) Reproduction planning (paper-specific)
- Problem definition, notation, assumptions, and objective functions.
- Algorithm steps and required components.
- Dataset generation/loading, training/optimization, and evaluation protocol.
- Baselines and ablations required for faithful reproduction.
5) Python implementation principles
- Separate concerns (data, models/algorithms, training/solver loop, evaluation, utils, config).
- Prefer low coupling and high cohesion.
- Split code into modules whenever responsibilities can be separated.
- Prefer one clear responsibility per file.
- Keep a single source file under ~200 lines whenever practical.
- If a file grows beyond ~200 lines, refactor into submodules unless there is a clear reason not to.
- deterministic seeds when applicable,
- explicit config for key hyperparameters,
- clear experiment entry points.
6) README header requirements (paper metadata)
- **Paper title** (original title as published)
- **Authors** (full names, affiliations, and email addresses if available)
- **Abstract** (verbatim copy of the original abstract)
- Paper title: provide Chinese translation if original is in English; keep original if paper is in Chinese.
- Authors: keep original names and affiliations; translate country/region names if needed.
- Abstract: provide faithful Chinese translation of the abstract.
```markdown
# [Paper Title]
**Authors:** Author Name¹, Co-Author Name²
**Affiliations:**
¹ Department, University, Country (email@university.edu)
² Lab, Institution, Country (email@institution.edu)
## Abstract
[Verbatim abstract text from the paper]
---
[Then reproduction project content begins...]
```
7) README update requirements (bilingual + images)
- `README.md` (English original)
- `README_zh-CN.md` (Chinese translation aligned with the English version)
- paper citation and target claims to reproduce,
- environment/setup commands,
- project structure overview and module responsibilities,
- how to run main experiments,
- expected outputs/metrics and where artifacts are saved,
- known deviations from the paper and why.
8) Output contract
- source-derived extraction notes/file(s) (TeX-first, PDF fallback when needed),
- implemented/updated Python code,
- `README.md` and `README_zh-CN.md` with embedded generated images.
- exact extracted content (verbatim from source),
- your implementation notes and engineering decisions.
- which claims/experiments were successfully reproduced,
- known gaps or deviations from paper results, with reasons.
More tools from the same signal band
Order food/drinks (点餐) on an Android device paired as an OpenClaw node. Uses in-app menu and cart; add goods, view cart, submit order (demo, no real payment).
Sign plugins, rotate agent credentials without losing identity, and publicly attest to plugin behavior with verifiable claims and authenticated transfers.
The philosophical layer for AI agents. Maps behavior to Spinoza's 48 affects, calculates persistence scores, and generates geometric self-reports. Give your...