LLM-Generated Skills Fail to Outperform Basic Prompting in Data Science Workflows

Ablation studies across 56 data science tasks show that LLM-generated skill files provide no reliable performance improvement over standard prompting. Evaluation across data preparation, extraction, statistical analysis, and reporting stages indicates that automated skill curation does not currently mitigate the manual bottleneck of expert-written guidance.