What problem does it solve? Running a machine learning experiment involves many error-prone manual steps: writing training code, deploying it to a GPU, monitoring long-running jobs, and collecting multi-seed results. This Skill automates that entire pipeline for experiments already planned in a research wiki, keeping experiment status and logs synchronized. ## Core Features & Use Cases - Three run modes: deploy mode launches a planned experiment and returns immediately, collect mode checks a running experiment and gathers results when finished, and full mode runs all four phases end-to-end for quick local jobs. - Local and remote deployment: run on a local GPU via screen sessions or deploy to a remote server over SSH with GPU selection, code sync, and result pull-back. - Automated code generation and review: generates train.py, config.yaml, run.sh, and requirements.txt from the experiment plan, with an optional LLM code review pass and a mandatory small-scale sanity check. - Use Case: You have an experiment page in wiki/experiments/ with status planned. Run the deploy mode to generate code, sanity-check it, and launch it on a free GPU; hours later, run collect mode to parse seed results, compute mean and standard deviation, and mark the experiment completed. ## Quick Start Run the experiment with slug my-exp by asking the assistant to execute /exp-run my-exp in deploy mode, then later collect results with /exp-run my-exp --collect.
