Create a workflow script for docs/ml-research-workflow-spec.md as rough guideline for ML research (+ add additional verifications/critics to the workflow where needed) and save it to .claude/workflows/ml-research.js so that I can run it later with the Workflow tool. The agents you use in the workflow should use the needed skills defined in ~/Development/extern/hf-skills. HF Jobs may fail for several reasons and error, and analysis and retry is important. Also, prior to the preflight HF job, train/eval the model locally on CPU with 1 step each to verify the code works correctly. The script should work with any ml research task a user may ask for.