Experiment planner
Skill maximussthegreat/ml-researcher-os/skills/experiment-planner
Agent skills and workflows for reproducible ML research.
npx -y skills add maximussthegreat/ml-researcher-os --skill experiment-plannerAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when turning an ML hypothesis, paper claim, or model idea into a reproducible experiment plan with baselines, metrics, seeds, risks, and expected artifacts.
SKILL.md
1.7 KB, as published. Nobody here has run it
Experiment Planner
Use this skill after claim extraction and before writing training code.
Goal
Convert a research claim into the smallest controlled experiment that can produce useful evidence.
Required inputs
- Hypothesis or paper claim
- Dataset or proposed dataset
- Target metric
- Compute budget
- Existing baseline, if any
If any input is missing, ask for it or mark it as unknown and make a conservative plan.
Required output
Write an experiment plan with:
- Hypothesis
- Minimum viable experiment
- Baselines
- Ablations
- Data split strategy
- Metrics
- Seed policy
- Failure modes
- Logging and artifacts
- Stop criteria
Baseline rules
Every plan needs at least one baseline. Prefer:
- simplest non-neural baseline
- standard library baseline
- previously reported baseline from the source paper
- ablation of the proposed method
Failure modes to check
- data leakage
- train/validation split mismatch
- metric mismatch
- hidden preprocessing fit on validation or test data
- seed sensitivity
- tiny test set
- weak baseline
- cherry-picked run
- missing negative result
Rules
- Do not plan an experiment without at least one meaningful baseline.
- Do not compare a heavily tuned model against an untuned baseline.
- Do not claim reproduction when the dataset, split, or metric differs from the source claim.
- Prefer three small controlled runs over one expensive ambiguous run.
Output style
Be concrete. Prefer a small experiment that can run today over an impressive plan that cannot be verified.