ARIS — Auto Research In Sleep
Let Claude Code run ML research overnight — idea discovery, GPU experiments, and a cross-model adversarial review loop with GPT-5.6-Sol that scored a real paper from 5.0/10 to 7.5/10 across 20+ autonomous experiments.
Last reviewed by the paperbanana team on Jul 13, 2026
Install
git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
bash Auto-claude-code-research-in-sleep/tools/install_aris.sh ~/your-projectWhat is ARIS — Auto Research In Sleep?
ARIS — Auto Research In Sleep (repository name Auto-claude-code-research-in-sleep) is a set of plain-Markdown Claude Code skills for autonomous ML research workflows, built around one core idea: cross-model adversarial review. Claude Code executes the research — surveying literature, running GPU experiments, drafting the manuscript — while an external model, GPT-5.6-Sol accessed via Codex MCP, acts as a critical reviewer. The README argues this deliberately avoids the "local minima" of self-play: a single model reviewing its own work tends to share its own blind spots, whereas two different models create an adversarial rather than self-consistent check, similar to how adversarial bandit settings are fundamentally harder to game than stochastic ones. An optional `reviewer: oracle-pro` flag routes the toughest reviews through GPT-5.5 Pro via the Oracle MCP for even stronger reasoning.
The skill set — around 80 composable skills, all plain SKILL.md files with no framework, database, or daemon required — chains into named workflows covering the full research lifecycle: `/idea-discovery` (literature survey → 8-12 brainstormed ideas → novelty check → GPU pilot experiments → ranked report), `/experiment-bridge` (turn a plan into deployed, running experiments), `/auto-review-loop` (a 4-round autonomous review-fix-re-review cycle that in one documented real run moved a paper’s reviewer score from 5.0/10 borderline-reject to 7.5/10 review-ready across 20+ GPU experiments run overnight, unattended), `/paper-writing` (narrative → outline → figures → LaTeX → PDF → auto-review, with anti-hallucination citations pulled from DBLP/CrossRef instead of LLM-generated references), `/rebuttal` (parses reviewer comments and drafts a character-limited response gated by three safety checks — no fabrication, no overpromise, full reviewer-concern coverage), `/resubmit-pipeline` (text-only port of a polished paper to a new venue), and `/paper-talk` (paper → Beamer/PPTX slides → speaker notes → assurance audits). A `/research-pipeline` command chains discovery through writing end to end, and `/research-wiki init` turns on a persistent knowledge base so papers, ideas, and failed experiments are remembered across sessions instead of re-discovered each run.
Because every skill is a portable SKILL.md, ARIS is not locked to Claude Code + OpenAI: the README documents native support for Codex CLI, Cursor, Trae, Antigravity, and GitHub Copilot CLI, plus alternative model combinations (GLM, MiniMax, Kimi, LongCat, DeepSeek) that need no Claude or OpenAI API key at all, including a free tier via ModelScope. On the integrity side, a newer `/integrity-forensics` skill runs the companion Anti-Autoresearch forensics sweep — nine auditor dimensions plus a GRIM/GRIMMER/statcheck numeric core — against your own paper before submission, so obvious fabrication or numeric-consistency problems surface before an actual reviewer finds them; the README is explicit that a clean sweep is recorded as "no new blocker" rather than an acquittal.
Core capabilities
Cross-model adversarial review loop
Claude Code executes while GPT-5.6-Sol (via Codex MCP, tiered xhigh/ultra reasoning) reviews adversarially rather than self-reviewing; `/auto-review-loop` iterates review→fix→re-review for up to 4 rounds unattended, with a documented real run moving 5.0/10 to 7.5/10 across 20+ GPU experiments.
End-to-end idea discovery and experiment automation
`/idea-discovery` surveys literature and brainstorms 8-12 novelty-checked ideas with GPU pilot experiments; `/experiment-bridge` turns an approved plan into deployed experiments (local, remote SSH, or rented on-demand via Vast.ai with `gpu: vast`), auto-writing scripts and folding results back into the paper.
Anti-hallucination paper writing
`/paper-writing` goes narrative → outline → figures → LaTeX → PDF → auto-review in one command, fetching real BibTeX from DBLP/CrossRef instead of LLM-generated citations to eliminate hallucinated references.
Safety-gated rebuttal drafting
`/rebuttal` parses reviewer comments and drafts a character-limited response that will not finalize unless three gates pass: no fabrication (every claim maps to paper/review/user-confirmed result), no overpromise (every promise user-approved), and full reviewer-concern coverage.
Persistent Research Wiki and self-optimization
`/research-wiki init` gives ARIS memory of papers, ideas, and failed experiments across sessions (failed ideas become anti-repetition memory); `/meta-optimize` analyzes usage logs to propose data-driven SKILL.md improvements to the toolkit itself.
Flexible model backends with no vendor lock-in
Every skill is a plain SKILL.md, so the same workflows run on Codex CLI, Cursor, Trae, Antigravity, or Copilot CLI, and support alternative reviewer/executor combinations (GLM, MiniMax, Kimi, LongCat, DeepSeek) requiring no Claude or OpenAI API key.
What you can use it for
Turning a research direction into a scored, revised paper overnight
Run `/research-pipeline "your research direction"` to chain idea discovery, experiments, the auto-review loop, and paper writing end to end while you sleep.
Improving an existing paper with its own codebase
Run `/research-pipeline "improve method X" — ref paper: <arxiv URL>, base repo: <github URL>` so ARIS reads the paper’s weaknesses, clones the codebase, and generates ideas that specifically fix those weaknesses with that code.
Drafting a safe, grounded rebuttal under a tight deadline
Run `/rebuttal "paper/ + reviews" — venue: ICML, character limit: 5000` to get a paste-ready response plus an extended draft, with GPT-5.6-Sol stress-testing it before it finalizes.
Preparing a conference talk after acceptance
Run `/paper-slides "paper/"` and `/paper-poster-html "paper/"` for a Beamer/PPTX deck with speaker notes and a measurement-gated, print-ready poster.
Auditing your own paper for integrity issues before submission
Run `/integrity-forensics`, which launches the Anti-Autoresearch sweep across nine auditor dimensions plus numeric consistency checks (GRIM/GRIMMER/statcheck) and blocks submission on unresolved obligation-bearing findings.
How to get started
- 1
Clone and install the skills
Run `git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git` then `bash Auto-claude-code-research-in-sleep/tools/install_aris.sh ~/your-project` to symlink ARIS skills into `<project>/.claude/skills/` (or `cp -r skills/* ~/.claude/skills/` for a global install).
- 2
Set up the Codex MCP reviewer
Install the Codex CLI (`npm install -g @openai/codex`), run `codex setup` and set the model to `gpt-5.6-sol`, then register it with Claude Code via `claude mcp add codex -s user -- codex mcp-server`.
- 3
Run a workflow inside Claude Code
Start `claude` and invoke a workflow directly, e.g. `/idea-discovery "your research direction"`, `/auto-review-loop "your paper topic"`, `/paper-writing "NARRATIVE_REPORT.md"`, or the full `/research-pipeline "your research direction"`.
- 4
Tune effort and enable memory
Append inline overrides like `— effort: max` or `— difficulty: nightmare` to any command for deeper/harder review passes, and run `/research-wiki init` once per project to turn on persistent cross-session memory.
How it compares to similar skills
ARIS is the heaviest-duty option here for experiment-running ML research specifically, not general literature or manuscript-only work. For non-experimental disciplines, a lighter suite fits better.
Academic Research Skills
Pick academic-research-skills if you need a general research-to-publication pipeline with citation-integrity gates but do not require GPU-experiment orchestration or cross-model review.
Auto Empirical Research Skills
Pick auto-empirical-research-skills if your empirical work is social-science econometrics (DiD, IV, RDD) rather than ML experiments needing GPU deployment and code review.
Medical Research Skills
Pick medical-research-skills if your research is clinical or biomedical rather than ML/CS experiment-driven — ARIS’s GPU experiment loop is built for deep-learning-style work.