What problem does it solve?
This Skill removes the gap between an experiment plan and actually running the experiments, so you can generate initial results without wasting GPU hours on incorrect implementations.
Core Features & Use Cases
- Implements experiments from a plan: Reads EXPERIMENT_PLAN.md (and related logs) to extract run order, hyperparameters, metrics, and success criteria, then generates the corresponding training/evaluation code.
- Deploys and collects initial results: Optionally clones a base repo, sanity-runs first, then deploys the full experiment suite with parallel routing and result collection.
- Prevents costly mistakes with review: Performs an optional cross-model code review (e.g., GPT-5.4 xhigh) before deployment to catch logic errors and evaluation-ground-truth mistakes.
Use Case Example: When you already have a claim-driven EXPERIMENT_PLAN.md from earlier workflows, this Skill converts it into runnable scripts, sanity-checks them, deploys the experiments, and updates refine-logs/EXPERIMENT_TRACKER.md with an initial results summary.
Quick Start
Ask to implement and execute the plan by providing the experiment plan path, for example: implement and run experiments from refine-logs/EXPERIMENT_PLAN.md.
