arXiv Review Paper Harness skill
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
- 466 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Python scripts write the paper project to disk and run pdflatex and bibtex
- The agent calls the arXiv API and opens web pages to verify citations
- The collaborating-with-claude and collaborating-with-gemini bridges launch other models from the command line and pass them pieces of the work
Install
- Works in
Manual install
git clone https://github.com/appautomaton/latex-arxiv-SKILL.gitThe skills live in .codex/skills of the repository. Open the cloned folder in Codex or copy the subfolders you need into your project's skills directory.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Python scripts write the paper project to disk and run pdflatex and bibtex
- The agent calls the arXiv API and opens web pages to verify citations
- The collaborating-with-claude and collaborating-with-gemini bridges launch other models from the command line and pass them pieces of the work
This is third-party code. Review the repository files before installing.
What it does
arXiv Review Paper Harness helps write a machine learning and AI review paper in LaTeX. Its main skill, arxiv-paper-writer, walks the agent through a fixed chain. The agent first takes a literature snapshot on the topic without writing any paper text, then scaffolds an IEEEtran project with a plan and candidate titles. Until you approve the plan and an issues.csv task table exists, not a paragraph may go into main.tex. Each section then becomes a tracked task with target citations and acceptance criteria. For every task the agent looks up sources, writes text and checks the citation against a live page before adding it to ref.bib. An unsupported claim becomes a TODO instead of an invented reference. After that the latex-rhythm-refiner skill evens out sentence and paragraph length while keeping citations intact. The work counts as done when pdflatex and bibtex build the project with no undefined citations. Deterministic steps live in Python scripts: scaffolding, the plan, CSV validation, compilation and arxiv_registry.py, which searches papers through the arXiv API and caches metadata and BibTeX in SQLite. The collaborating-with-claude and collaborating-with-gemini bridges let the agent ask another model for a second opinion.
Who it is for. For researchers and graduate students writing an ML or AI review in LaTeX who want every reference in it verified.
Good fit when
- You need a state-of-the-field review, for example of generative image models, of 6–10 pages in the IEEEtran template
- You want to approve the structure and title before the agent starts writing prose
- An existing LaTeX project needs its ref.bib audited and repaired without re-scaffolding
- The draft reads at a monotone rhythm and needs varied sentences while every citation stays intact
Not a fit when
- You need a new experimental paper: SKILL.md lists it as a case where the skill does not apply, and the README allows it only with a tailored plan
- You need non-academic documents: the skill is tuned for review papers
- The machine has neither pdflatex with bibtex nor latexmk: without a build the completion criterion cannot be checked
- The agent has no web access: checking citations against a live source does not work without it
Example request
Write a review article for arXiv about state-of-the-art generative image models. Start with a literature snapshot and a plan, and begin the prose only after I approveLimitations
- It needs pdflatex with bibtex or latexmk, Python 3.8 or newer and web search for the agent to check citations.
- The author tested the set on macOS with GPT-5.2 Extra High. The repository has no separate test records for Claude Code.
- The arxiv_registry.py script calls the arXiv API, and the collaborating-with-claude and collaborating-with-gemini bridges launch other models from the command line.
- The skill is built for reviews. For original research the plan has to be adapted by hand.
How to disable. Delete the arxiv-paper-writer, latex-rhythm-refiner and both bridge folders from the agent's skills directory. Paper projects it created live in the folder you passed with --out and stay in place.
FAQ
How does the skill guard against invented references?
Every citation is checked against a live source before it enters ref.bib. A claim without support becomes a TODO. At the end pdflatex and bibtex must report no undefined citations.
Why does the agent not start writing right away?
AGENTS.md allows only headings, bullets and seed citations until the plan is approved and the task table exists. That keeps the text from sprawling out of control.
How many prompts does it take to get a paper?
In the example/v0-single-SKILL run two prompts produced the paper: the first asks for a review on the topic, the second hands the choice of title and scope to the agent. The repository holds two finished papers with plans, task tables, sources and PDFs.
Can I use the skill on an existing LaTeX project?
Yes, auditing and repairing ref.bib runs without a new scaffold: just give the agent the project path.
Does the set work in Claude Code?
The README claims dual compatibility with Codex and Claude Code because the skills follow the Agent Skills standard. The author tested the set in Codex, so check the output in Claude Code on a small topic first.
Discussion
No questions or reviews yet.
Related
SkillsEight skills for finished papers and theses in LaTeX and Typst: compile, references, language polish, cover letters and a reviewer-style audit
The paper-spine skill takes a topic, materials and data through a scientific paper: literature, text, figures, review and Word, LaTeX and PDF files
Skill that adapts a research manuscript to a journal: it builds writing rules from published papers and revises a manuscript by section with a log
A ready-to-fork Claude Code template for academic work: slides, papers and data analysis with multi-agent review and quality gates
A skill for writing and revising ML, CV and NLP papers: section structure, paragraph flow and reviewer-style self-review
Claude Code skills for the academic pipeline: research, paper drafting, peer review and citation checks
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| arXiv Review Paper Harness (this entry) | no | 466 | |
| PaperOrchestra | no | 684 | a multi-agent pipeline from plan and figures to literature review, without mandatory approval gates |
| Academic Paper Skills | no | 1,416 | two skills, a strategist and a composer, with quality checks and no LaTeX project build |
README in short
The README opens with a flowchart of the process: literature snapshot, scaffold, approval, task table, writing loop, rhythm pass, build. Then come a table of benefits with hard quality gates, a two-prompt quickstart, the contents of the set, two example papers, requirements, a short FAQ and the MIT license.
README badge
Original SKILL.md
--- name: arxiv-paper-writer description: > Write LaTeX ML/AI review articles for arXiv using the IEEEtran template and verified BibTeX citations. Use when writing, planning, or continuing an arXiv review/survey paper, or when validating and repairing citations in an existing LaTeX project. compatibility: > Python 3.8+ for scripts. Web browsing/search for citation verification. LaTeX is required (pdflatex + bibtex or latexmk). metadata: short-description: ML/AI review papers (IEEEtran template) with verified citations --- # ML/AI Review Paper Workflow (IEEEtran template) ## When to Use - ML/AI review papers for arXiv (main text ~6-10 pages; references excluded) - LaTeX + BibTeX workflow with verified citations - Citation validation/repair on existing LaTeX projects ## When NOT to Use - Novel experimental research papers (this is a review workflow) - Non-academic documents ## Inputs - Topic description (required) - Constraints: venue, page limit, author/affiliations (optional) - Existing project path for citation validation (optional) ## Outputs - `main.tex` (LaTeX source) - `ref.bib` (verified BibTeX entries) - `IEEEtran.cls` - `plan/<timestamp>-<slug>.md`, `issues/<timestamp>-<slug>.csv` - Figures/tables; `main.pdf` - `notes/research-snapshot.md` (Gate 0 research snapshot) - `notes/literature-notes.md` (optional per-citation notes) - `notes/arxiv-registry.sqlite3` (arXiv metadata/BibTeX cache) **Conventions**: run `python3 scripts/...` from this skill folder (where `scripts/` lives); `<paper_dir>` is the paper/project root (contains `main.tex`, `ref.bib`, `plan/`, `issues/`, `notes/`). Paths like `plan/...` are under `<paper_dir>`. Papers are created under `--out <workspace_dir>` (your working directory, never inside this skill folder); pass the same `--out` (and `--name`, if used) to both bootstrap stages. For arXiv discovery/metadata/BibTeX, use `scripts/arxiv_registry.py` (no ad-hoc curl/wget). --- ## Gated Workflow