auto-research-with-eyes

AutoResearchWithEyes

Let Claude Code do research while you sleep. Wake up to find your paper scored, weaknesses identified, experiments run, and narrative rewritten — autonomously.

· Join Community

A Claude Code plugin for autonomous ML research workflows. Orchestrates cross-model collaboration — Claude Code drives the research while an external LLM (via Codex MCP) acts as a critical reviewer. Also supports alternative model combinations (e.g., GLM + GPT, GLM + MiniMax) — no Claude API required.

Why cross-model? A single model reviewing its own output creates blind spots. Two complementary models — Claude Code for fast execution, GPT-5.4 xhigh for rigorous critique — produce better outcomes than either alone. Going from 1 to 2 models is the biggest gain; adding more gives diminishing returns.

---

Quick Start

# 1. Clone and enter the project
git clone https://github.com/llv23/AutoResearchWithEyes.git
cd AutoResearchWithEyes

# 2. Set up Codex MCP (for cross-model review)
npm install -g @openai/codex
codex auth login
# Codex MCP auto-configures from .mcp.json when running in the project directory

# 3. Launch Claude Code — skills and commands are auto-discovered
claude

# Run any workflow command:
> /autor.idea-discovery "your research direction"      # Plain text → ./autor.idea_discovery/
> /autor.idea-discovery path/to/autor.idea_discovery/task_spec.md    # Spec file  → path/to/autor.idea_discovery/
> /autor.auto-review-loop                              # Review → fix → re-review overnight
> /autor.paper-writing "NARRATIVE_REPORT.md"           # Narrative → polished PDF
> /autor.research-pipeline "your research direction"   # Full end-to-end pipeline

Each command produces a self-contained output folder — the input spec, intermediate results, final reports, and live status tracking are all organized in one place:

path/to/
└── autor.idea_discovery/                  # Self-contained output folder
    ├── task_spec.md                       # Copy of input (portability & reproducibility)
    ├── task_status.md                     # Live status, decisions, linked idea rankings
    ├── LITERATURE_SURVEY.md               # Phase 1: landscape summary with paper table
    ├── IDEA_REPORT.md                     # Phase 5: ranked ideas + experiment plan
    └── REVIEW_<topic>_<date>.md           # Phase 4: external reviewer feedback

task_status.md is updated after each phase — it tracks pipeline progress, key decisions, and includes a ranked idea table where each idea links to its full description in IDEA_REPORT.md.

See click-through/case-0/ for a complete worked example.

Output Folder Conventions

All workflow commands follow the same self-contained output pattern: the output folder includes task_spec.md (input), task_status.md (live status with linked rankings), and all generated results — making it portable, shareable, and reproducible.

Command	Input	Output Folder
/autor.idea-discovery path/to/spec.md	spec.md	path/to/autor.idea_discovery/
/autor.idea-discovery "plain text"	(inline)	./autor.idea_discovery/
/autor.idea-discovery spec.md — output: custom/	spec.md	custom/

Override the output folder with — output: path/to/output/ appended to any command.

See Setup for full details.

Features

• 10 composable skills — atomic building blocks: literature search, idea generation, novelty check, experiments, paper writing
• 4 workflow commands — orchestrate skills + agents into end-to-end pipelines (/autor.idea-discovery, /autor.auto-review-loop, /autor.paper-writing, /autor.research-pipeline)
• 2 specialized agents — research-reviewer (senior ML reviewer via Codex MCP) and paper-improver (2-round auto-improvement)
• Cross-model collaboration — Claude Code executes, GPT-5.4 xhigh reviews. Adversarial, not self-play
• Centralized configuration — all constants in CLAUDE.md, override per-invocation with inline arguments
• Venue templates — bundled template directories with fallback resolution: TEMPLATE_DIR/VENUE/ → bundled → error with instructions
• GPU deployment — auto rsync, screen sessions, multi-GPU parallel experiments, live monitoring
• Flexible models — default Claude x GPT-5.4, also supports GLM + GPT, GLM + MiniMax

---

Architecture