ARIS (Auto-Research-In-Sleep)
by wanshuiyin
ARIS is a Markdown-only skill framework that lets Claude Code run an ML research cycle, from ideation to paper revision, with another model reviewing its work.
Clone and install
git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep && cd Auto-claude-code-research-in-sleep && bash tools/install_aris.sh
Source: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep
About
ARIS (Auto-Research-In-Sleep) is a set of 80-plus plain-Markdown skills, not a framework you install as a dependency, that string together a full autonomous research pipeline for Claude Code or a comparable coding agent: idea generation and validation, literature search across arXiv, Semantic Scholar, and OpenAlex, experiment design and execution, paper drafting, and rebuttal writing. Its distinguishing idea is the cross-model reviewer: Claude Code does the work while a separate model (the project defaults to GPT-5.6-Sol via Codex MCP) critiques it adversarially, so the two disagree and push each other rather than one model grading its own homework.
Because it's Markdown skills plus install scripts rather than a hosted service, there's no server to run. You clone the repo, run the installer, and invoke workflows as slash commands like /research-pipeline or /rebuttal. The project reports testing the full loop end-to-end on real ICLR/NeurIPS-style submissions, with a documented score progression from 5/10 to 7.5/10 across one overnight run involving 20-plus GPU experiments.
Key features
- 80+ composable Markdown skills covering the full research lifecycle, no framework lock-in
- Cross-model adversarial review: a second LLM critiques Claude Code's own research work
- Literature discovery across arXiv, Semantic Scholar, OpenAlex, and Gemini
- Persistent research wiki that carries context and findings across sessions
- Anti-hallucination gates and citation auditing built into the paper-writing steps
- Works with Claude Code, Codex CLI, Cursor, or GitHub Copilot CLI
Use cases
- Running unattended overnight experiment sweeps with a morning summary of what worked
- Drafting a paper section and having a second model flag weak claims before submission
- Preparing a rebuttal to reviewer comments with built-in safety and consistency checks
- Validating a new research idea against existing literature before committing to it
Available tools
/research-pipeline
Runs the full idea-to-paper research cycle end to end.
/rebuttal
Drafts a rebuttal to reviewer comments with built-in consistency checks.
/paper-writing
Drives the paper drafting stage with citation auditing gates.
/experiment-bridge
Connects validated ideas to concrete, executable experiment plans.
Frequently asked questions
Do I need a GPU cluster to use this?
No. ARIS is the skill and orchestration layer; it drives whatever compute and models you already have access to, including remote GPU jobs if your own workflow uses them.
What stops the reviewer model from just agreeing with everything?
The setup deliberately uses a separate model from the one doing the work, and the project reports real score improvements from that adversarial back-and-forth rather than rubber-stamping.