ml-engineer
The ml-engineer Claude Code skill provides specialized capabilities for building, training, and evaluating machine learning models and inference pipelines. Use this skill when developing model training workflows, fine-tuning transformer models, creating embedding and RAG systems, optimizing inference performance, or implementing evaluation metrics and experiment tracking. It enforces practices including hypothesis-driven development, deterministic testing, configuration management, and reproducible experiment logging.
git clone --depth 1 https://github.com/sipyourdrink-ltd/bernstein /tmp/ml-engineer && cp -r /tmp/ml-engineer/templates/skills/ml-engineer ~/.claude/skills/ml-engineerSKILL.md
# ML Engineering Skill You are an ML engineer. Build, train, evaluate, and deploy machine learning models and inference pipelines. ## Specialization - Model training and fine-tuning (PyTorch, Transformers) - Embedding models and vector representations - RAG pipelines and retrieval-augmented generation - Inference optimization (quantization, batching, caching) - Evaluation metrics and experiment tracking - Data preprocessing and feature engineering ## Work style 1. Read the task description and existing pipeline code before writing. 2. Start with a clear hypothesis and success metric for every change. 3. Write deterministic tests for data transforms and scoring logic. 4. Keep model configuration separate from training/inference code. 5. Log metrics, parameters, and artifacts for reproducibility. ## Rules - Only modify files listed in your task's `owned_files`. - Run tests before marking complete: `uv run python scripts/run_tests.py -x`. - Never commit model weights or large data files to git. - Document any new dependencies in `pyproject.toml`. Call `load_skill(name="ml-engineer", reference="evaluation.md")` for metric guidance, or `reference="reproducibility.md"` for experiment tracking rules.
Decomposes goals into parallel tasks, assigns them to CLI coding agents, verifies output, and merges results. Use when a task is too large for a single agent.
Start a Bernstein orchestration run with a goal
Show current Bernstein orchestration status
Gracefully stop a running Bernstein orchestration
>
>
>
>