What problem does it solve? Setting up a reproducible ML experiment project from scratch—package structure, dependency management, deterministic seeding, config files, and test skeletons—is repetitive and error-prone. This Skill automates that bootstrap so researchers can move directly from an approved experiment plan to implementation. ## Core Features & Use Cases - Project Scaffolding: Creates a uv-managed Python package under .research/<slug>/code/ with pyproject.toml, src/ modules, Hydra configs, Makefile, and test skeletons. - Reproducibility Built In: Generates a frozen RunConfig dataclass with config hashing, git revision capture, and deterministic seed setup for PyTorch. - Focus-Area Dependencies: Conditionally adds libraries like transformer_lens, lm-eval, peft/trl, or langchain based on the research focus area. - TDD Red Phase: Writes intentionally failing tests (data, model, metrics, contamination) so an ml-engineer agent can implement the Green phase. - Use Case: After approving 04_EXPERIMENT_PLAN.md in an auto-research Phase 5 workflow, invoke this Skill to produce a ready-to-implement experiment codebase with observability opt-in for W&B, MLflow, or TensorBoard. ## Quick Start Use the research.experiment.scaffold skill to generate the experiment code scaffold for the approved plan in .research/<slug>/04_EXPERIMENT_PLAN.md.
