It is tempting to fix a wrong answer by rewriting the prompt. Sometimes that works. More often it papers over a data problem that will resurface the moment the input shifts.
If the same question returns different answers on different days, the issue is rarely the wording. It is the ground truth underneath.
Fix the source, not the sentence
- Clean the data the agent reads from before you tune how it reads.
- Make one source authoritative, and point everything at it.
- Track where answers come from, so you can trace a wrong one back.
- Prefer fewer, trusted sources over many noisy ones.
Why this pays off
A system grounded in consistent data degrades gracefully. When something changes, you know where to look. A system held together by clever prompts breaks quietly, and you find out from a user.
Clever prompting is a loan against your future debugging time.