← ClaudeAtlas

exp-reviewlisted

Review experiment code and Jupyter notebooks on two axes in parallel sub-agents: standards (does the code follow the repo's documented coding standards?) and quality (reproducibility, fixed seeds, data leakage, security, scalability, error handling). Use when the user wants to review an experiment, notebook or ML pipeline before moving on.
samuhs/mad-science · ★ 3 · Code & Development · score 74
Install: claude install-skill samuhs/mad-science
Check whether the code follows the quality rules. ### 1. Reproducibility * **Clean Execution:** Can the code run end-to-end without failing if the kernel is restarted? * **Dependency Management:** Are the libraries and exact versions documented (e.g., `requirements.txt`, `Pipfile`)? * **Artifact & Data Versioning:** Are data paths and model artifacts versioned (e.g., DVC, MLflow, immutable S3 URIs) rather than relying on local, hardcoded paths (`C:/user/data.csv`)? * **Fixed Seeds:** Is the `random_state` or `seed` fixed for all stochastic processes (data splits, weight initializations, ML algorithms) to guarantee deterministic results? ### 2. Security * **Credential Leakage:** Are there any hardcoded passwords, API keys, access tokens, or DB URIs? Instruct the user to move these to environment variables or secret managers. * **Sensitive Data (PII):** Is the model exposing or directly logging Personally Identifiable Information (e.g., SSN, names, emails)? Ensure data is anonymized or hashed. * **Query Sanitization:** If the code dynamically queries databases, is it protected against SQL injection? ### 3. Scalability * **Memory Optimization:** Are data types optimized (e.g., using `float32` instead of `float64`, or `category` instead of `object` in Pandas)? Will the code break if data volume doubles? * **Vectorization:** Flag any iterative loops (like `for` loops or `iterrows`) over datasets. Suggest vectorized operations using NumPy/Pandas. * **Framework Suitability:** Do