08-14-2026, 02:02 PM
Data work fails when one model is asked to do everything in a single prompt. Break it into stages and pass outputs forward:
1. Clean – "Fix missing values and inconsistent types in this data; list what you changed"
2. Transform – "Aggregate this by week, output a table"
3. Analyze – "Find the top 3 patterns in this table, with numbers"
4. Summarize – "One-paragraph executive summary for a non-technical reader"
Each stage gets clean context and a single job – errors show up where they happen instead of hiding in a 4-page output.
Bonus rule: ask the model to restate the numbers ("show your math") so you can verify.
What's your chain for data tasks – one mega-prompt or staged?
1. Clean – "Fix missing values and inconsistent types in this data; list what you changed"
2. Transform – "Aggregate this by week, output a table"
3. Analyze – "Find the top 3 patterns in this table, with numbers"
4. Summarize – "One-paragraph executive summary for a non-technical reader"
Each stage gets clean context and a single job – errors show up where they happen instead of hiding in a 4-page output.
Bonus rule: ask the model to restate the numbers ("show your math") so you can verify.
What's your chain for data tasks – one mega-prompt or staged?