Run ML Research Overnight with ARIS's 87 Claude Code Skills

An agent that "does research" usually means one long chat that forgets its own dead ends. ARIS (Auto-claude-code-research-in-sleep), an 18k-star MIT-licensed repo by wanshuiyin, turns the research lifecycle into 87 composable Claude Code skills and puts a second model family in the reviewer's chair.

Why This Skill Matters

ARIS calls itself a methodology, not a platform. The skills cover literature and novelty search, idea discovery, GPU experiments, an autonomous review loop, paper writing and peer review, and they chain into named pipelines: /idea-discovery (Workflow 1), /experiment-bridge (1.5), /auto-review-loop (2), /paper-writing (3), /rebuttal (4), /resubmit-pipeline (5), /paper-talk (6), or /research-pipeline for Workflows 1 through 3 end to end.

The design choice that runs through all of them is cross-model adversarial review: Claude executes, GPT-6-Astra at xhigh effort reviews, and GPT-5.5 Pro through Oracle is an optional third seat. Citations are checked against DBLP and CrossRef, and an optional Research Wiki keeps papers, ideas, failed experiments and claim status across sessions — failed ideas become what the README calls anti-repetition memory.

Installation

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-project

That symlinks the skills into <project>/.claude/skills/. For a global copy instead, cp -r Auto-claude-code-research-in-sleep/skills/* ~/.claude/skills/; for a subset, --list-groups, --groups X,Y or --skills X. Updating later is git pull followed by bash tools/smart_update.sh --apply, which updates safe skills and flags your own customizations.

The review skills call GPT through a Codex MCP server that ARIS ships itself:

npm install -g @openai/codex && codex login
claude mcp add codex -s user -- python3 "$(pwd)/Auto-claude-code-research-in-sleep/mcp-servers/codex-exec/server.py"

Restart Claude Code and confirm claude mcp list shows codex connected. The README flags why this matters: codex-cli 0.154 removed codex mcp-server, the entry point older ARIS installs registered, so anyone set up before 2026-09-11 has to claude mcp remove codex -s user and re-add it against mcp-servers/codex-exec/server.py. Cursor, Trae, Antigravity and Copilot CLI use the same codex key, with adaptation guides under docs/; Codex CLI has its own managed install via tools/install_aris_codex.sh.

Real Workflow: From a Direction to a Ranked Idea Report

1. Start claude in your project and give Workflow 1 a direction. The README's own warning: be specific — not "NLP" but "factorized gap in discrete diffusion LMs".

/idea-discovery "factorized gap in discrete diffusion LMs"

2. The skill surveys recent papers, open problems and recurring limitations, then brainstorms 8–12 concrete ideas through GPT-6-Astra xhigh.
3. Ideas are filtered by feasibility, compute cost and a quick novelty search; the top ones go through a deep novelty check plus a devil's-advocate review.
4. The top 2–3 are piloted in parallel on different GPUs, 30 minutes to 2 hours each, and ranked by empirical signal — ideas with positive pilot results rise.
5. The top idea is refined into a problem-anchored proposal through iterative GPT-6-Astra review, and a claim-driven experiment plan is written with ablations, budgets and run order.

You get a ranked IDEA_REPORT.md, plus refine-logs/FINAL_PROPOSAL.md and refine-logs/EXPERIMENT_PLAN.md for the top idea. Dead-end ideas are documented too. Already have a plan? /experiment-bridge implements, deploys and collects results. Already have a paper? Hand Workflow 2 to the night:

/auto-review-loop "your paper topic or scope"

GPT-6-Astra reviews, Claude fixes — rewriting sections, adding baselines, or running new experiments via /run-experiment, while skipping and flagging anything estimated over 4 GPU-hours — then re-evaluates and repeats until the score reaches POSITIVE_THRESHOLD (default 6/10) or MAX_ROUNDS (default 4). If the context window fills mid-loop, it resumes from REVIEW_STATE.json. The README's real run: 5.0/10 to 7.5/10 across 4 rounds and 20+ GPU experiments.

Tips

• Targeted mode takes a paper and a codebase in one line: /research-pipeline "improve method X" — ref paper: <arXiv URL>, base repo: <GitHub URL>. Paper only asks "what can be improved?"; repo only asks "what can I build with this code?".
• /research-wiki init is one command. After it, /research-lit ingests papers automatically, /idea-creator reads the wiki before brainstorming and writes ideas back, and /result-to-claim updates claim status.
• /rebuttal "paper/ + reviews" — venue: ICML, character limit: 5000 will not finalize unless three gates pass: every claim maps to the paper, a review or a user-confirmed result; every promise is user-approved; every reviewer concern is tracked. The character limit is required. You get PASTE_READY.txt at the exact count and REBUTTAL_DRAFT_rich.md for manual editing.
• Workflow 2 needs your GPU server configured in CLAUDE.md, or gpu: vast to rent from Vast.ai.
• No Claude or OpenAI API? The README documents 10 alternative routes in docs/MODEL_COMBINATIONS.md, including Z.ai GLM, Kimi, Qwen, MiniMax, DeepSeek-V3.1 via ModelScope, OpenRouter as a reviewer backend, Codex as executor with Claude or Gemini reviewers, and Google Antigravity.

When Not to Use This

ARIS is built around two things you may not have. The first is a GPU: Workflow 2's review-driven experiments deploy to a server you configure in CLAUDE.md (or rent via Vast.ai), and without one the loop has nowhere to run the ablations the reviewer asks for. The second is a second model: the reviewer seat is the point, the README's default is GPT-6-Astra through a Codex login, and the alternative routes above exist for when you cannot use it. If you want a single-model setup with no MCP server, you can still install individual skills, but the cross-model review the README headlines is not what you will get. And the scope is research — papers, experiments, rebuttals, talks, patents — not general-purpose writing.

See the leaderboard for more skills.