LaTeX Agent SKILL (arXiv ML/AI Review Papers)

English · 简体中文

[!NOTE] This repo implements a .codex SKILL: an end-to-end, issues-driven workflow that scaffolds an IEEEtran two-column LaTeX project for ML/AI review articles on arXiv.org, with verified BibTeX citations. Experimental Claude Code support is provided via .claude → .codex.

What’s in here

• End-to-end workflow: issues-driven paper writing (plan generation → user approval → issues CSV generation → per-issue writing → rhythm refinement → QA/compile).
• Primary SKILL: arxiv-paper-writer at .codex/skills/arxiv-paper-writer/SKILL.md.
• Helper scripts: scaffolding, planning, validation, compilation under .codex/skills/arxiv-paper-writer/scripts/.
• Example output: a generated paper in example/ (including example/main.pdf).

[!NOTE] The workflow is inspired by “issue-driven development” as demonstrated by the appautomaton/agent-designer project.

Quickstart (2 prompts → a compiled example paper)

The example/v0-single-SKILL paper in this repo was generated by activating the arxiv-paper-writer SKILL with two user prompts.

Prompt 1 (start the paper)

write a review article for arxiv that is about SOTA generative image models

After this first prompt, the agent:

• does an initial literature pass,
• drafts a section framework,
• proposes a few candidate titles,
• and writes a plan/<timestamp>-<slug>.md file with clarification questions and a draft plan.

[!TIP] Open the generated plan/ markdown file and answer the clarification questions to steer scope, title, and coverage.

Prompt 2 (delegate decisions and proceed)

I will let you choose the best title and with the topics and inclusion of material that you see the best fit

In this example, the second prompt is intentionally vague (and the plan questions were ignored). The agent still makes best-effort choices and generates a complete LaTeX project that compiles to a PDF.

[!NOTE] See the result in example/main.tex, example/ref.bib, and example/main.pdf.

Environment notes

[!IMPORTANT] A working LaTeX environment is required (e.g., pdflatex + bibtex, or latexmk).

• Tested on macOS with GPT-5.2 (Extra High).