paper-reading-skill
paper-reading-skill turns one arXiv AI or robotics paper into a report with each claim tied to evidence in the source.
- 148 ★ on the repository
- Last commit
Minimal access
Why this level
- The skill downloads pages and PDFs and runs local scripts in the working directory.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/paper-reading-skillDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill downloads pages and PDFs and runs local scripts in the working directory.
This is third-party code. Review the repository files before installing.
What it does
The skill for Codex and other agents reads one arXiv paper and writes a self-contained Markdown report. Scripts first resolve the latest version, download the PDF and web page, and extract references and figures. The agent then classifies the work as theory, method, system, dataset, empirical study or survey. The type decides what counts as evidence: a theorem and counterexample, an ablation, a protocol, dataset composition. Every key claim lands in a claim, evidence and verdict table, and the deciding passage in the original is located in an appendix. There are three modes: quick read, full reviewer-level review, and review with a reproducibility assessment.
Who it is for. Researchers and engineers who read AI and robotics papers and want to know whether the authors' conclusions hold up.
Good fit when
- Before adopting a method you need to know whether the paper supports its claimed effect.
- A reading group wants a report with quotes and tables taken from the original.
- You must judge whether results can be reproduced with your own resources.
- A theory paper without experiments should be read through its proofs, not a template.
Not a fit when
- You want a short retelling of the abstract.
- You need a survey of dozens of papers at once.
- The source is a local PDF, a DOI link or an ordinary website: the pipeline needs an arXiv ID.
Example request
Use $paper-reading to read this paper: https://arxiv.org/abs/2510.12796Limitations
Works with a single paper that has an arXiv ID. No license is declared in the repository, so check the terms before building it into a workflow. Before the first run `scripts/bootstrap.sh` creates a virtual environment and installs `requirements.txt`. The report is written by a model: verify conclusions and quotes against the original. Latest commit in the repository: August 2026. Downloading and processing go over the network, so access to arXiv and the sites hosting paper versions must be open. The skill targets AI and robotics papers; for other fields the template needs manual checking.
How to disable. Delete the skill directory from the agent skills folder. Working directories named `{arxiv_id}_{title}` with reports and cache are removed separately.
FAQ
What comes out?
One Markdown file `{arxiv_id}_阅读报告.md` in the `{arxiv_id}_{title}` folder. `raw/`, `images/`, `cache/` and `logs/` hold sources and the run log beside it.
How do I run preprocessing by hand?
Run `bash scripts/run_pipeline.sh "2510.12796"` in the skill root. Both an arXiv link and a bare ID work.
What language is the report in?
Templates and examples in the repository are Chinese and the report is named «阅读报告». You can ask for another language, but check the result.
Is a check needed after writing?
Yes. The validator `scripts/validate_report.py` checks that all key claims and deciding evidence reached the report.
Which paper types does the skill handle?
Theory, methods, systems and platforms, datasets and benchmarks, empirical studies, surveys and opinion pieces. Each type gets its own evidence carriers.
Discussion
No questions or reviews yet.
Related
SkillsTwo skills for Claude Code and Codex: the agent translates papers in the EasyRead library and answers your margin notes
Takes an agent from a topic to an arXiv review paper in LaTeX IEEEtran: approved plan, every citation checked against a live source, PDF build
Eight skills for LaTeX research papers, from topic to compiled PDF, with verified BibTeX citations and gated steps
A skill set for academic papers: deep method analysis, visual slides and a coherent write-up from a paper PDF
A Claude Code skill for research papers: a five-stage pipeline, section rules, a style audit and an independent red-team review.
A skill that checks a manuscript against practices of 883 ICLR, ICML and NeurIPS Oral papers and suggests edits
README in short
The template fixes the questions put to a paper and who owes what evidence, but does not demand formulas, tables or figures in every work. If the original has no figures, none are invented.