Key idea: A prompt is a set of ingredients, and you can measure each one. Spelling out the format, defining your categories, and covering edge cases usually matter far more than telling the model who it is.
Prompt ingredients
Prompt preview · 82 characters, exactly what the model gets
Task
Tag this customer feedback with a theme, a sentiment, and whether it's actionable.
- Run with every ingredient off. How many answers are even valid JSON?
- Turn on Output format only and run again. Every answer is now usable; how many are right?
- Add Category definitions. Watch the theme column.
- Turn on Role. Does "You are a senior PM…" change the score?
- Try 1 example, then 3. Which tricky item do the examples fix?
- Add Edge-case rules. Which item could only the rules fix?
- With everything on, switch to Smart, then turn ingredients off one at a time. Does the bigger model still need them?