experiment-design（实验设计）

Experiment Design

Design structured, progressive experiment plans for research papers.

Input

• $0 — Research idea, plan, or method description

References

• 4-stage progressive experiment prompts: ~/.claude/skills/experiment-design/references/stage-prompts.md

Scripts

Generate experiment design

python ~/.claude/skills/experiment-design/scripts/design_experiments.py --plan research_plan.json --output experiment_design.json
python ~/.claude/skills/experiment-design/scripts/design_experiments.py --method "contrastive learning" --task classification --format markdown

Generates baselines, ablation matrix, hyperparameter grid, metric selection. Stdlib-only.

4-Stage Progressive Framework (from AI-Scientist-v2)

Stage 1: Initial Implementation

• Focus on getting a basic working implementation
• Use a simple dataset
• Aim for basic functional correctness
• Completion: at least one working (non-buggy) implementation

Stage 2: Baseline Tuning

• Tune hyperparameters (learning rate, epochs, batch size)
• Do NOT change model architecture
• Test on at least TWO datasets
• Completion: stable training curves, improvement over Stage 1

Stage 3: Creative Research

• Explore novel improvements and insights
• Be creative and think outside the box
• Test on at least THREE datasets
• Completion: demonstrated novel improvement

Stage 4: Ablation Studies

• Systematic component analysis
• Each ablation tests a different aspect
• Use same datasets as Stage 3
• Completion: all planned ablations done

Output Format

{
  "stages": [
    {
      "name": "initial_implementation",
      "goals": ["Basic working baseline", "Simple dataset"],
      "max_iterations": 5,
      "completion_criteria": "Working implementation with non-zero accuracy"
    }
  ],
  "baselines": ["Method A", "Method B"],
  "datasets": ["Dataset1", "Dataset2", "Dataset3"],
  "metrics": ["accuracy", "F1", "inference_time"],
  "ablation_components": ["component_A", "component_B"],
  "hyperparameter_grid": {
    "lr": [1e-4, 1e-3, 1e-2],
    "batch_size": [32, 64, 128]
  },
  "num_seeds": 3
}

Rules

• Always start simple (Stage 1) before complex experiments
• Each stage builds on the best result from the previous stage
• Multi-seed evaluation for statistical significance
• Document every experiment run in notes.txt
• Generate figures for training curves and comparisons

Related Skills

• Upstream: research-planning, idea-generation
• Downstream: experiment-code, data-analysis
• See also: paper-assembly