A good data work prompt isn't a magic trick. It's a checklist you don't have to remember, run by something that never gets bored.
Keep a negative list
Write down what you never want as explicitly as what you do. The exclusions often do more work than the instructions.
- Name the audience and the single goal before you generate
- Compare two outputs instead of trusting one
- Have the model flag what it could not cover
Reuse the recipe
Treat the first output as a draft, not a verdict. The prompt does the typing; you keep the judgment about what ships.
Start with the smallest version
Ask for the reasoning, then the result. When the model shows its steps you can see exactly where to push back.
Treat the first output as a draft, not a verdict. The prompt does the typing; you keep the judgment about what ships.