appautomaton/latex-arxiv-SKILL

A highly customizable agentic harness for arXiv-ready ML/AI review papers (and beyond). It drives agentic AI like Codex CLI and Claude Code through a gated LaTeX workflow with verified BibTeX citations.

What it solves

It automates the creation of academic review papers for AI and machine learning, specifically targeting the arXiv format. It solves the problem of "hallucinated" citations and poor structural quality common in AI-generated text by enforcing a disciplined, gated pipeline that requires human approval and strict verification of every reference before it is included in the final document.

How it works

The project acts as an agentic harness (a set of rules and scripts) for coding agents like Claude Code or OpenAI Codex. It follows a specific sequence:

  1. Research & Planning: The agent identifies relevant papers and creates a draft plan and project scaffold.
  2. Approval Gate: The user must approve the plan before any prose is written.
  3. Issue-Driven Writing: The project uses a CSV file as an "execution contract." Each section of the paper is treated as a tracked issue with specific acceptance criteria.
  4. Verification: Every citation is web-checked against a live source before being added to the BibTeX file.
  5. Refinement & Compilation: A dedicated "rhythm refiner" improves prose flow, and the system ensures the project compiles via pdflatex and bibtex without errors or undefined citations.

Who it’s for

Researchers and AI practitioners who need to produce high-quality, professionally formatted (IEEEtran) review or survey papers without manually managing every citation and structural detail.

Highlights

  • Zero-Hallucination Citations: Every \cite{} is verified against a live source.
  • Issue-Driven Workflow: Progress is tracked via a CSV file rather than vague prompts, ensuring no untracked work.
  • Portable Agent Skills: Compatible with both Anthropic Claude Code and OpenAI Codex.
  • Deterministic Tooling: Uses Python scripts for scaffolding, arXiv discovery (with SQLite caching), and LaTeX compilation to ensure consistent results.