aris-autonomous-research

---
name: aris-autonomous-research
description: ARIS (Auto-Research-In-Sleep) — Markdown-only autonomous ML research workflows using cross-model review loops, idea discovery, experiment automation, and paper writing with Claude Code or any LLM agent.
triggers:
  - run autonomous research pipeline
  - set up ARIS research workflow
  - use claude code for ML research
  - automate paper writing with AI
  - cross-model research review loop
  - run experiment automation with ARIS
  - install ARIS skills for claude code
  - generate research ideas while sleeping
---

# ARIS — Autonomous Research In Sleep

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

ARIS is a **zero-dependency, Markdown-only** autonomous ML research system. Each "skill" is a plain `SKILL.md` file that any LLM agent can read and execute. The system orchestrates **cross-model collaboration**: one model executes (Claude Code / Codex) while another critiques (GPT-5.4 / Gemini / GLM / MiniMax), breaking self-review blind spots without any framework or lock-in.

Core capabilities:
- 🔬 **Idea discovery** from a research direction or existing paper
- 🧪 **Experiment automation** with GPU-ready code generation and W&B tracking
- 📝 **Paper writing** (LaTeX, Beamer slides, A0 poster)
- 🔁 **Cross-model review loops** with score progression
- 📬 **Rebuttal drafting** with safety gates (no fabrication, no overpromise, full coverage)

---

## Installation

### 1. Clone the repository

```bash
git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
cd Auto-claude-code-research-in-sleep

2. Install skills into Claude Code

Copy the skills directory to your project, or symlink it:

# Option A: copy skills to your project
cp -r skills/ /your/project/.claude/skills/

# Option B: symlink (keeps skills up to date)
ln -s /path/to/Auto-claude-code-research-in-sleep/skills /your/project/.claude/skills

Claude Code auto-discovers SKILL.md files in .claude/skills/**. No registration step needed.

3. Configure the MCP reviewer (cross-model review)

ARIS uses the llm-chat MCP server so the executor model can call a second model for review. Install it:

cd mcp-servers/llm-chat
pip install -r requirements.txt   # or: uv pip install -r requirements.txt

Add to your claude_desktop_config.json (or Claude Code MCP config):

{
  "mcpServers": {
    "llm-chat": {
      "command": "python",
      "args": ["/path/to/Auto-claude-code-research-in-sleep/mcp-servers/llm-chat/server.py"],
      "env": {
        "OPENAI_API_KEY": "$OPENAI_API_KEY",
        "LLM_MODEL": "gpt-4o"
      }
    }
  }
}

For alternative reviewers (Kimi, GLM, MiniMax, DeepSeek) set LLM_BASE_URL and LLM_MODEL to the provider's OpenAI-compatible endpoint. No Claude or OpenAI API required.

4. (Optional) Codex MCP for OpenAI Codex as executor

{
  "mcpServers": {
    "codex": {
      "command": "npx",
      "args": ["@openai/codex-mcp"],
      "env": {
        "OPENAI_API_KEY": "$OPENAI_API_KEY"
      }
    }
  }
}

---

Environment Variables

Variable	Required	Description
OPENAI_API_KEY	For GPT reviewer	OpenAI API key
ANTHROPIC_API_KEY	For Claude executor	Anthropic API key
LLM_BASE_URL	Alternative reviewer	OpenAI-compatible base URL
LLM_MODEL	Alternative reviewer	Model name at that endpoint
WANDB_API_KEY	Experiment tracking	Weights & Biases key

---

Workflows & Commands

Full pipeline (idea → paper)

/research-pipeline "factorized gap in discrete diffusion LMs"

With a reference paper and base codebase:

/research-pipeline "improve positional encoding in transformers" — ref paper: https://arxiv.org/abs/2104.09864, base repo: https://github.com/facebookresearch/fairseq

Parameters:

Flag	Default	Effect
ref paper	—	ARIS reads this paper, finds weaknesses, targets them
base repo	—	Clone and use this repo as experiment base
compact: true	false	Generate lean summary files (good for short-context models)

---

Workflow 1 — Idea Discovery

/idea-discovery "sparse attention in long-context LLMs"

What it does:

1. Searches recent arXiv papers in the direction
2. Identifies open gaps and contradiction clusters
3. Generates 3–5 novel ideas with novelty scores
4. Runs /research-refine to sharpen the best idea into a problem statement

---

Workflow 1.5 — Experiment Bridge

/experiment-bridge "idea_proposal.md" — base repo: https://github.com/huggingface/diffusers

What it does:

1. Reads the sharpened idea from Workflow 1
2. Generates GPU-ready experiment code
3. Runs GPT cross-model code review before deployment (code review: true by default)
4. Executes training loop with W&B logging
5. Saves results to experiment_results/

Example generated experiment scaffold:

# experiment_results/run_001/train.py  (auto-generated by /experiment-bridge)
import wandb
import torch
from torch.utils.data import DataLoader

wandb.init(
    project=os.environ.get("WANDB_PROJECT", "aris-experiment"),
    config={
        "method": "factorized_discrete_diffusion",
        "lr": 3e-4,
        "epochs": 50,
        "batch_size": 32,
    }
)

for epoch in range(config.epochs):
    for batch in dataloader:
        loss = model(batch)
        wandb.log({"loss": loss.item(), "epoch": epoch})

---

Workflow 2 — Literature Review

/literature-review "discrete diffusion language models"

Anti-hallucination: ARIS verifies every citation via DBLP → CrossRef → marks unverified as [VERIFY]. Never fabricates BibTeX.

---

Workflow 3 — Paper Writing

/paper-write "experiment_results/ + idea_proposal.md" — venue: NeurIPS

Supported venue templates: CVPR, NeurIPS, ICML, ICLR, ACL, AAAI, ACM MM

---

Workflow 4 — Rebuttal

/rebuttal "paper/ + reviews/" — venue: ICML, character limit: 5000

Parameters:

Parameter	Default	Description
venue	ICML	Target conference
character limit	Required	Hard character limit
quick mode	false	Stop after strategy (Phase 0–3), no draft
auto experiment	false	Auto-run experiments when reviewers ask for new evidence
max stress test rounds	1	GPT-5.4 stress-test passes on draft
max followup rounds	3	Per-reviewer follow-up round limit

Three safety gates — rebuttal will NOT finalize if any fails:

• 🔒 No fabrication — every claim maps to paper/review/confirmed result
• 🔒 No overpromise — every promise is user-approved
• 🔒 Full coverage — every reviewer concern is tracked

Outputs:

• PASTE_READY.txt — exact character count, ready to paste to venue portal
• REBUTTAL_DRAFT_rich.md — extended version for manual editing

---

Presentation & Poster

/paper-slides "paper/"     # Beamer PDF + PPTX + speaker notes + Q&A prep
/paper-poster "paper/"     # A0/A1 poster PDF + PPTX + SVG (venue colors)

---

Standalone Utility Skills

Skill	Command	What it does
training-check	/training-check "train.py"	Diagnose training instability, NaN, slow convergence
result-to-claim	/result-to-claim "results.json"	Convert raw numbers into paper-ready claims
ablation-planner	/ablation-planner "idea.md"	Design minimal ablation study for a method
experiment-plan	/experiment-plan "idea.md"	Claim-driven experiment roadmap
research-refine	/research-refine "idea.md"	Sharpen vague idea into problem-anchored proposal
formula-derivation	/formula-derivation "method.md"	Develop and verify research formulas
paper-illustration	/paper-illustration "paper/"	Generate figures (Gemini-assisted)
grant-proposal	/grant-proposal "idea.md"	Draft grant proposal from research idea

---

Alternative Model Combinations

ARIS requires only an OpenAI-compatible endpoint for the reviewer. Set environment variables:

# Kimi as reviewer
export LLM_BASE_URL="https://api.moonshot.cn/v1"
export LLM_MODEL="moonshot-v1-128k"
export LLM_API_KEY=$MOONSHOT_API_KEY

# DeepSeek as reviewer
export LLM_BASE_URL="https://api.deepseek.com/v1"
export LLM_MODEL="deepseek-chat"
export LLM_API_KEY=$DEEPSEEK_API_KEY

# MiniMax as reviewer
export LLM_BASE_URL="https://api.minimax.chat/v1"
export LLM_MODEL="abab6.5s-chat"
export LLM_API_KEY=$MINIMAX_API_KEY

Then in mcp-servers/llm-chat/server.py the LLM_BASE_URL env var overrides the OpenAI default. No code changes needed.

---

Using with Codex CLI (no Claude)

ARIS ships a parallel skills-codex/ directory with the same workflows adapted for OpenAI Codex CLI:

# Install Codex CLI
npm install -g @openai/codex

# Run a workflow
codex --skill skills/skills-codex/research-pipeline/SKILL.md \
  "improve contrastive learning in vision transformers"

---

Using with Cursor

1. Open Cursor settings → Rules → paste content of docs/CURSOR_ADAPTATION.md
2. Copy skills/ to .cursorrules-skills/ in your project
3. In chat: @research-pipeline "your research direction"

---

Using with Trae (ByteDance IDE)

See docs/TRAE_ARIS_RUNBOOK_EN.md for full setup. Trae supports SKILL.md natively via its plugin system.

---

Input Templates

Pre-filled templates for every workflow live in templates/:

templates/
  research-pipeline.md      # Full pipeline input
  idea-discovery.md
  experiment-bridge.md
  literature-review.md
  paper-write.md
  rebuttal.md
  paper-slides.md
  paper-poster.md

Use a template:

/research-pipeline — template: templates/research-pipeline.md

---

Project Structure

Auto-claude-code-research-in-sleep/
├── skills/
│   ├── research-pipeline/SKILL.md      # Main orchestration workflow
│   ├── idea-discovery/SKILL.md         # Workflow 1
│   ├── experiment-bridge/SKILL.md      # Workflow 1.5
│   ├── literature-review/SKILL.md      # Workflow 2
│   ├── paper-write/SKILL.md            # Workflow 3
│   ├── rebuttal/SKILL.md               # Workflow 4
│   ├── paper-slides/SKILL.md
│   ├── paper-poster/SKILL.md
│   ├── training-check/SKILL.md
│   ├── result-to-claim/SKILL.md
│   ├── ablation-planner/SKILL.md
│   ├── experiment-plan/SKILL.md
│   ├── research-refine/SKILL.md
│   ├── formula-derivation/SKILL.md
│   └── skills-codex/                   # Codex CLI variants
├── mcp-servers/
│   └── llm-chat/                       # OpenAI-compatible reviewer MCP
├── templates/                          # Input templates per workflow
├── docs/
│   ├── CURSOR_ADAPTATION.md
│   ├── TRAE_ARIS_RUNBOOK_EN.md
│   ├── ANTIGRAVITY_ADAPTATION.md
│   ├── MODELSCOPE_GUIDE.md             # Free tier setup
│   ├── MiniMax-GLM-Configuration.md
│   └── CODEX_GEMINI_REVIEW_GUIDE.md
└── README.md

---

Common Patterns

Pattern 1: Start from an arXiv paper you want to beat

/research-pipeline "improve method" — ref paper: https://arxiv.org/abs/2406.04329, base repo: https://github.com/org/repo

ARIS reads the paper → identifies weaknesses → clones repo → generates targeted ideas → runs experiments → writes paper.

Pattern 2: Resume interrupted session

Add compact: true to any workflow. ARIS writes a lean SESSION_SUMMARY.md. On resume:

/research-pipeline — resume: SESSION_SUMMARY.md

Pattern 3: Jump into the middle of a pipeline

Already have results? Jump to paper writing:

/paper-write "my_results/ + my_idea.md" — venue: NeurIPS

Already have a paper? Jump to rebuttal:

/rebuttal "paper/ + reviews/" — venue: ICML, character limit: 5000

Pattern 4: Free tier via ModelScope

export LLM_BASE_URL="https://api-inference.modelscope.cn/v1"
export LLM_MODEL="Qwen/Qwen2.5-72B-Instruct"
export LLM_API_KEY=$MODELSCOPE_API_KEY

See docs/MODELSCOPE_GUIDE.md for zero-cost setup.

---

Troubleshooting

Skills not discovered by Claude Code

Ensure SKILL.md files are under .claude/skills/ relative to your project root. Claude Code scans this path at startup.

MCP reviewer not connecting

# Test the llm-chat server directly
cd mcp-servers/llm-chat
python server.py --test
# Should print: {"status": "ok", "model": "gpt-4o"}

W&B logging not working in experiment-bridge

wandb login  # uses WANDB_API_KEY env var, or prompts for manual entry

Citation hallucination in literature-review

All unverified citations are tagged [VERIFY] in output. Search DBLP manually for flagged entries before including in your paper. Never remove the [VERIFY] tag without confirming.

Rebuttal exceeds character limit

ARIS tracks character count per section. If a draft exceeds the limit, it automatically trims supporting evidence (keeps claims, removes elaboration). You can also pass quick mode: true to get the strategy without the draft, then write targeted sections manually.

Cross-model review loop not running (self-review fallback)

If the llm-chat MCP is unreachable, ARIS falls back to single-model review with a warning in the output. Check MCP server logs:

tail -f ~/.claude/mcp-logs/llm-chat.log

Session context overflow

Use compact: true on any workflow invocation to produce a compressed SESSION_SUMMARY.md that fits in a fresh context window.

---

Extending ARIS

Every skill is a plain Markdown file. To create a custom skill:

# my-custom-skill

## Trigger
When the user says "run my custom analysis"...

## Steps
1. Read input files
2. Call `mcp__llm-chat__chat` with the review prompt
3. Write output to `custom_output/`

## Output
- `custom_output/analysis.md`
- `custom_output/score.json`

Save as .claude/skills/my-custom-skill/SKILL.md and Claude Code will discover it automatically.