The trap with writing prompts is asking for the finished thing. Ask for options instead, then bring your taste to the edit.
Make the output reviewable
Lock the parts that never change and leave one slot for the variable. Now every run shares a family resemblance instead of starting from scratch.
- Lower the reading level, then check what broke
- Save the prompt that worked so the next person reuses it
- Have the model flag what it could not cover
Reuse the recipe
Write down what you never want as explicitly as what you do. The exclusions often do more work than the instructions.
- Ask for three angles, then cut to the one worth shipping
- Keep one ask per prompt — chaining beats cramming
- Compare two outputs instead of trusting one
Ask for the reasoning, then the result. When the model shows its steps you can see exactly where to push back.