arXiv Review Paper Harness
An agentic harness, packaged as a portable Agent Skill for Claude Code and Codex. It turns a topic into an arXiv-ready ML/AI review paper that is gated, issue-driven, and verified citation by citation.
arXiv Review Paper Harness is an agentic harness for writing machine-learning and AI review papers in LaTeX. You give your coding agent a topic, and the harness drives it through a disciplined pipeline: literature discovery, a human approval gate, an issue-by-issue writing loop, citation verification, prose refinement, and compilation. The result is a two-column IEEEtran project that compiles to a PDF. Its skills and scripts follow the portable Agent Skills standard, so the same files run in OpenAI Codex and Anthropic Claude Code.
How it works
flowchart TD
A([Topic]) --> B[Research snapshot<br/>10 to 20 papers, no prose]
B --> C[Scaffold IEEEtran project<br/>plus draft plan and candidate titles]
C --> D{Approval gate}
D -- revise --> C
D -- approved --> E[Issues CSV<br/><i>execution contract</i>]
E --> F[Per-issue writing loop<br/>research, write, <b>verify every citation</b>]
F --> G[Rhythm refinement<br/>vary sentence and paragraph length]
G --> H[QA and compile<br/>fix overfull boxes, zero undefined cites]
H --> I([main.pdf and ref.bib])
style D fill:#f5b400,stroke:#333,color:#000
style I fill:#3da639,stroke:#333,color:#fff
The agent cannot write a single paragraph into main.tex until two conditions hold. First, you approve the plan. Second, the issues CSV exists. From there, every section is a tracked issue with target citations and acceptance criteria, and every citation is checked against a live source before it enters ref.bib.
Why it's different
- Hard quality gates, not vibes: no prose before plan approval, and nothing marked
DONEuntil acceptance criteria are met. - Verified citations only: every
\cite{}is web-checked against its source before being added, so there are no hallucinated references. - Issue-driven execution: an
issues/*.csvfile is the single source of truth for progress, and the agent splits or inserts issues as scope grows instead of doing untracked work. - Deterministic where it matters: scaffolding, plan and issue generation, arXiv discovery (with a local SQLite cache), validation, and compilation are Python scripts, so they behave the same on every run.
- Dual-runtime: one skill bundle that runs in both Claude Code and Codex.
- Compiles or it is not done: delivery requires a clean
pdflatexandbibtexbuild with no undefined-citation warnings.
Quickstart
The example/v0-single-SKILL paper was generated by activating the arxiv-paper-writer skill with two prompts.
Prompt 1 (start the paper):
write a review article for arxiv that is about SOTA generative image models
The agent does an initial literature pass, drafts a section framework, proposes candidate titles, and writes a plan/<timestamp>-<slug>.md with clarification questions.
[!TIP] Open the generated
plan/file and answer the clarification questions to steer scope, title, and coverage.
Prompt 2 (delegate the decisions and proceed):
I will let you choose the best title and the topics and inclusion of material that you see the best fit
This second prompt is intentionally vague, and the plan questions were ignored. The harness still makes best-effort choices and produces a complete, compiling LaTeX project. See main.tex, ref.bib, and main.pdf.
What's in the box
| Component | What it is |
|---|---|
arxiv-paper-writer | The primary harness. The gated workflow, guardrails, and success criteria. |
scripts/ | Deterministic Python helpers: scaffolding, plan and issue generation, CSV validation, LaTeX compile, and arxiv_registry.py (arXiv Atom-API discovery and BibTeX with a local SQLite cache). |
latex-rhythm-refiner | Post-processes prose for readable sentence and paragraph rhythm while preserving every citation. |
collaborating-with-claude · -gemini | Bridges to delegate sub-tasks or get a second opinion from another model. |
agent-skills-standard.md | The repo's spec for authoring portable SKILL.md bundles across Codex and Claude Code. |
example/ | Two fully generated papers, with plans, issue CSVs, sources, and compiled PDFs. |
Examples
| Paper | Citations | Notes |
|---|---|---|
| v0: Generative image models review | 55 verified | Single-skill run, the 2-prompt quickstart above. |
| v0.5: Video world simulators (3D/4D) review | 81 verified | Multi-skill run with the SQLite arXiv registry and BibTeX cache. |
Requirements
[!IMPORTANT] A working LaTeX environment is required:
pdflatexandbibtex, orlatexmk.
- Agent runtime: OpenAI Codex or Anthropic Claude Code, with skills enabled.
- Python 3.8+ for the helper scripts.
- Web search and browsing for citation verification.
- Tested on macOS with GPT-5.2 (Extra High).
FAQ
How does it prevent hallucinated or invented citations?
Guardrails, built into the workflow. Every citation is verified against a live source before it enters ref.bib, and any claim without evidence becomes a TODO rather than a fabricated reference.
Can I use it on an existing LaTeX project?
Yes. Point it at your project and a citation-validation pass audits and repairs ref.bib without re-scaffolding anything.
Can it write original or experimental research papers, not just reviews? Yes, with a little tailoring. Review and survey articles are its sweet spot out of the box, but nothing locks it there. Shape the plan and inputs to your goal, and the same gated workflow extends to original or experimental work.
Credits
The issue-driven workflow is inspired by "issue-driven development" as demonstrated by appautomaton/agent-designer.
License
MIT.