latex-paper-skills

New: Beginner Workflow Guide (Chinese)

English | 简体中文

Portable AI agent skill bundle for writing ML/AI academic papers — from topic to compiled PDF, with verified BibTeX citations, gated workflows, and multi-agent collaboration.

What This Repo Is

This repository is not a single paper project. It is a reusable skill bundle for AI-assisted academic writing.

In this public snapshot, the repo keeps:

• the reusable writing skills under .codex/skills/
• the repo-level workflow and documentation
• two curated showcase projects: one empirical and one review

The repo intentionally does not keep bulky local datasets, machine-local settings, or cache databases. For the empirical showcase, the raw dataset used during development has been removed, but the retained code, result files, figures, and final PDF still show that the workflow completed a real data-and-experiment loop.

Status & Tested Models

The showcases in this repo were developed and tested with GPT-5.2 xhigh. Using GPT-5.4 is expected to yield better results. Test coverage is still limited — the workflow has significant untapped potential, though undiscovered bugs may exist.

How It Works

Optional multi-agent collaboration:

Skill Catalog

All skills live under .codex/skills/. Each has a SKILL.md (the executable spec an agent follows), plus optional scripts/, assets/, and references/.

Core Pipeline

Collaboration Layer

Gated Workflow

Both writer skills enforce strict gates:

Quick Start

Prerequisites

• Python 3.8+
• LaTeX environment (pdflatex + bibtex, or latexmk)
• AI agent runtime that reads SKILL.md (Codex CLI, Claude Code, etc.)

Full pipeline (topic → PDF)

Use the paper-from-zero skill. My topic is: <your topic>

Direct review paper

Use the arxiv-paper-writer skill. Write a review article about <topic>.

Direct empirical paper

Use the empirical-paper-writer skill. Write an experimental paper about <topic>.

Prompt templates: specify datasets and cloud execution

You can specify datasets, runtime budget, and an intended cloud platform directly in natural language.

Use the empirical-paper-writer skill.
Topic: <topic>.
Mandatory datasets: <Dataset A>, <Dataset B>.
Primary dataset: <Dataset A>.
Do not use private data.
Design a compute-aware experiment plan for a single A100 80GB.

Use the empirical-paper-writer skill.
Topic: <topic>.
Dataset: <dataset name> from local path <path>.
Target runtime: cloud A100 x1, max 8 hours.
Need a local smoke run first, then a full cloud run.
Keep the claims bounded by this compute budget.

Use the empirical-paper-writer skill.
Topic: <topic>.
Mandatory datasets: <dataset names>.
Target platform: AutoDL / Lambda / Slurm cluster.
Generate experiments/, configs/default.yaml, and an experiments README for cloud execution.
Assume I will run the jobs myself and then use results-backfill after real CSVs are available.

Notes:

• The skill can design for your dataset and cloud budget.
• The skill does not automatically provision or submit cloud jobs for you.
• Real results still need to be produced outside the AI session and written to paper/results/.

Shared Script Engine

.codex/skills/arxiv-paper-writer/scripts/ contains the core tooling shared by both writers:

Shared utilities (paper_utils.py, source_policy_utils.py) live in .codex/skills/_shared/.

Non-Negotiable Rules

• No prose before approval — main.tex stays skeleton-only until the plan is approved and the issues CSV exists.
• Issues CSV is the contract — update status per issue; only mark DONE when acceptance criteria are met.
• Citations must be verified — every citation is checked against an online source before entering ref.bib.
• Never fabricate citations, results, or significance claims.

Public Showcases

This public snapshot keeps two curated showcase projects:

If you only want the fastest tour:

1. Open projects/rt-inflow-forecast-closed-loop/README.md for the empirical route.
2. Open projects/peft-survey-2022-2026/README.md for the review route.
3. Read BEGINNER_WORKFLOW.zh-CN.md if you want the end-to-end workflow explanation in Chinese.

Project Structure

latex-paper-skills/
├── .codex/skills/
│   ├── paper-from-zero/              # Router: topic → writer skill
│   ├── arxiv-paper-writer/           # Review paper executor + shared scripts
│   ├── empirical-paper-writer/       # Empirical paper executor
│   ├── results-backfill/             # Back-fill real results into draft
│   ├── latex-rhythm-refiner/         # Prose polisher
│   ├── collaborating-with-claude/    # Claude Code bridge
│   ├── collaborating-with-gemini/    # Gemini CLI bridge
│   ├── check-collaborators/          # CLI health check
│   ├── _shared/                      # Shared utilities across skills
│   └── _orchestration/               # Workflow orchestration config
├── projects/
│   ├── rt-inflow-forecast-closed-loop/  # Empirical showcase
│   └── peft-survey-2022-2026/           # Review showcase
├── picture/                          # SVG diagrams for README
├── ARCHITECTURE.md                   # Detailed architecture analysis
├── BEGINNER_WORKFLOW.zh-CN.md        # Beginner workflow guide (Chinese)
├── README.md
└── README.zh-CN.md

Related

• Forked from renocrypt/latex-arxiv-SKILL
• Workflow inspired by appautomaton/agent-designer (issue-driven development)
• Uses IEEEtran LaTeX class

Star History

License

This repository is released under the MIT License.

Community

• LinuxDO