How to describe what you want so the AI builds the right thing
Describe the outcome and the rules around it, not the implementation. Say who uses it, what they should be able to do, what must never happen, and what it should look like when it has worked. Most disappointing results come from an instruction that was ambiguous rather than one that was too short.
The four things worth including
- Who is using this, and what they are trying to get done.
- What should happen in the ordinary case, in one sentence.
- What must never happen. Constraints do more work than features.
- How you will know it worked, in terms you could check yourself.
One thing at a time
A request holding five changes gets you five half-changes and no way to tell which one broke something. Ask for one, check it, then ask for the next. This is the single biggest difference between people who get good results and people who do not, and it costs nothing when building is not metered.
Say what happened, not what to do
When something is wrong, describe the symptom precisely, including what you did and what you saw. "The date shows as yesterday for anybody outside the UK" is worth more than an instruction to fix time zones, because it lets the agent find the real cause instead of the one you guessed at.