idea2code

Turn a research idea into running code. One command, one experiment.

"Talk is cheap. Show me the code." But now in the era of AI coding, the reverse might be true — your ability to create a new idea is far more important than implementing it.

This tool bridges that gap. You bring the idea. It writes the code.

/idea-iter try attention gates in the decoder

What it does

You describe an idea. The agent:

• Finds relevant papers to inform the implementation
• Reads your codebase to understand the architecture
• Discusses the plan with you before writing any code
• Makes surgical code edits to implement your idea
• Commits, pushes, and launches the experiment
• Returns immediately so you can start the next idea

Install

Option A: npx (requires Node.js)

npx idea2code

Option B: git clone (no Node.js needed)

git clone https://github.com/haoyudong-97/idea2code.git /tmp/idea2code && \
  rm -rf ~/.claude/skills/idea-iter ~/.claude/skills/check-experiments ~/.claude/skills/combine-findings ~/.claude/skills/auto-loop ~/.claude/skills/reset-iterations ~/.claude/skills/reset-iterations && \
  cp -r /tmp/idea2code/skill/idea-iter ~/.claude/skills/ && \
  cp -r /tmp/idea2code/skill/check-experiments ~/.claude/skills/ && \
  cp -r /tmp/idea2code/skill/combine-findings ~/.claude/skills/ && \
  cp -r /tmp/idea2code/skill/auto-loop ~/.claude/skills/ && \
  cp -r /tmp/idea2code/skill/reset-iterations ~/.claude/skills/ && \
  for s in idea-iter check-experiments combine-findings auto-loop reset-iterations; do \
    cp -r /tmp/idea2code/skill/research_agent ~/.claude/skills/$s/; \
  done && \
  rm -rf /tmp/idea2code && \
  echo "Done! Skills installed."

Requirements

• Claude Code installed
• Python 3.10+
• Git

Uninstall

rm -rf ~/.claude/skills/idea-iter ~/.claude/skills/check-experiments ~/.claude/skills/combine-findings ~/.claude/skills/auto-loop ~/.claude/skills/reset-iterations

Skills

Command	What it does
/idea-iter <idea>	Implement one idea → papers → discuss → code → launch experiment
/idea-iter --auto <idea>	Same but skips confirmation — launches directly
/check-experiments	Check running experiments, collect results, suggest next steps
/combine-findings <input>	Integrate a paper URL, rough idea, or literature into current work
/auto-loop <goal>	Run multiple iterations automatically toward a high-level goal
/reset-iterations [reason]	Archive current state and restart from iter/1 (new dataset, new direction)

/idea-iter

The core skill. Give it a specific idea or a vague direction:

/idea-iter add attention gates to decoder skip connections    # specific → skips paper search
/idea-iter improve model generalization                       # exploratory → searches papers first

It always discusses the plan with you before implementing (unless --auto).

/auto-loop

Hands-free mode. Give a high-level goal, and it runs repeated iterations:

/auto-loop improve segmentation on small organs

It asks you upfront:

1. How many GPUs? (each runs a different iteration in parallel)
2. Stop by time or iterations? ("run for 12 hours" or "run 5 iterations")
3. Any constraints? ("only try attention-based methods")

Then it loops: formulate ideas → implement → launch → wait → collect results → formulate next ideas. All iterations stay focused on your goal. Running experiments are never killed — when the limit is reached, it waits for them to finish before reporting.

How It Works

Your idea
    ↓
Classify: specific or exploratory?
    ↓
(exploratory only) Find relevant papers — arXiv API + WebSearch
    ↓
Discuss plan with you
    ↓
Implement the idea (surgical code edits via Agent)
    ↓
Commit to git branch (iter/1-attention-gates)
    ↓
Launch experiment (GPU-aware, local or remote SSH)
    ↓
Return immediately — start next idea

state.json and progress.md track all iterations, metrics, and results. Each iteration records its method, results, and learnings.

License

MIT