ARIS：让AI在你睡觉时做科研——自动化研究流程的Markdown工作流

ARIS (Auto Research In Sleep) is a lightweight, zero-dependency Markdown research workflow that leverages cross-model collaboration (Claude Code + GPT-5.4) to automate the entire research process—from literature调研 and experiment design to paper writing and overnight iterative improvement. It supports multiple models and platforms, and has already helped produce papers accepted at top conferences. The core goal is to let researchers focus on creative thinking while AI handles executional tasks.

Academic research involves time-consuming steps like literature review, experiment design, and paper writing. ARIS addresses this with three core principles:

1. Zero dependency: Pure Markdown files, no complex databases or Docker setups.
2. Cross-model compatibility: Works with Claude Code, GPT-5.4, GLM, MiniMax, and more.
3. Adversarial review: Uses an executor-reviewer dual-model setup to avoid self-review blind spots, similar to having an opponent point out weaknesses.

ARIS uses a dual-model collaboration mechanism:

• Why two models?: Single-model self-review leads to local optima; dual models (e.g., Claude Code for execution + GPT-5.4 for review) create adversarial checks to find deeper issues.

Key workflows:

1. Idea Discovery: Literature scan → brainstorming → feasibility screening → novelty check → pilot experiments → method refinement.
2. Experiment Bridge: Convert plans to scripts → code review → small-scale validation → GPU deployment.
3. Auto Review Loop: Overnight iteration (review → suggest improvements → implement → deploy → rewrite) until quality达标.
4. Paper Writing: Outline → figure generation → LaTeX writing → auto improvement loops.

ARIS has delivered tangible results:

• A CS conference paper using full ARIS流程 scored 8/10, with reviewers noting 'distinct empirical findings and solid support'.
• An AAAI 2026 paper (by @xinbo820-web) using ARIS-Codex skills was accepted.
• A machine learning project improved from 5.0/10 to 7.5/10 via 4 rounds of overnight auto iteration (20+ GPU experiments, no human intervention), reaching submission standards.

ARIS has a flexible ecosystem:

• 31 skills: Core skills (research-lit, idea-creator) + community-contributed skills (grant-proposal, paper-poster).
• Multi-platform support: Works with OpenClaw, Cursor IDE, Trae (ByteDance IDE), Antigravity (Google IDE), and Codex CLI.
• Alternative model combinations: Default (Claude Code + GPT-5.4),国产方案 (MiniMax-M2.7 + GLM-5), free (ModelScope with 2000 daily calls).

Quality control features:

• Anti-hallucination citations: Uses DBLP/CrossRef for real BibTeX entries.
• Code review: GPT-5.4 checks code before GPU deployment to save resources.
• 合理性 check: Runs small experiments first to validate code.

Limitations:

• Requires GPU server access.
• Complex math推导 needs human verification.
• Result interpretation depends on domain knowledge.

Responsible use: ARIS is a tool—researchers must review ideas, question assumptions, and retain final decision-making.

Quick Start:

1. Install: git clone https://github.com/Unimposing-electroscope363/Auto-claude-code-research-in-sleep.git
2. Configure Codex MCP: npm install -g @openai/codex → codex setup
3. Use workflows: /idea-discovery, /experiment-bridge, /auto-review-loop, /research-pipeline.

Future Outlook: ARIS aims to make AI a research partner, lowering entry barriers for scientific discovery. It’s open-source under MIT license, welcoming community contributions.