Data Pipeline Manager

Make deliberate, documented data decisions before training. This skill prevents the most common experimental validity failures: undocumented preprocessing, leaky splits, contaminated test sets, and irreproducible data pipelines.

Use this skill when:

• deciding how to acquire, access, or download a dataset
• designing or auditing train/val/test split protocols
• preprocessing choices need documentation before they affect results
• data quality issues could corrupt baselines or ablations
• contamination between training and evaluation data is a reviewer risk
• dataset versions, seeds, and pipeline steps need to be pinned for reproducibility
• a paper claim depends on a specific data protocol and the protocol is not yet written down

Do not use this skill for running experiments — use run-experiment once the pipeline is validated. Do not use this skill to write result sections — use experiment-story-writer after results exist.

Pair this skill with: