What problem does it solve? Turning a written experiment plan into running GPU experiments requires writing training and evaluation code, catching bugs before burning GPU hours, and orchestrating deployment — a slow, error-prone manual process. ## Core Features & Use Cases - Plan-Driven Implementation: Parses EXPERIMENT_PLAN.md and EXPERIMENT_TRACKER.md to generate training, evaluation, and data-loading scripts with fixed seeds and JSON/CSV result output. - Cross-Model Code Review: Sends experiment code to GPT-5.5 (via Codex MCP) for correctness review before deployment, catching logic bugs like using model outputs as ground truth. - Sanity-First Deployment with Auto-Debug: Runs the smallest sanity experiment first, auto-diagnoses failures (OOM, CUDA errors, NaN divergence) with up to 2 patches plus 2 clean reimplements, then routes small batches to /run-experiment or large sweeps to /experiment-queue. - Use Case: After finishing idea discovery, you have refine-logs/EXPERIMENT_PLAN.md with 12 experiments across 4 milestones. Run this skill to implement all scripts, get them reviewed, deploy to your GPUs, and collect initial results ready for /auto-review-loop. ## Quick Start Ask the AI to implement and deploy the experiments from refine-logs/EXPERIMENT_PLAN.md using the experiment-bridge skill.
