How do you improve a prompt without tailoring it to a few known examples?
Review note: The answer refers to the original author's CV or target job offer — adapt it to your own experience.
Answer out loud or in your head first, then compare.
Answer
First I prepare examples and criteria for a good answer. I check which errors repeat, then change the instruction to address the specific problem. I keep some examples for the final evaluation and don't use them for continuous prompt tweaking.
If I keep adding exceptions for a few known cases, the result can look great only on those. I also need new, hard and ambiguous inputs. Clear-cut conditions, e.g. “the amount can't be negative”, are better checked in code. Instead of asking for the model's hidden chain of thought, I ask for the result, a short reason and evidence that can be verified.
Why it matters: Makes it easier to defend the prompt work in your CV.
Possible follow-ups: How do you notice contradictory rules in a long instruction?
How did it go? Be honest — “Not yet” puts the question on your review list.
Start with
- F-P01 · What is an LLM? — LLM fundamentals and AI engineering
- F-P03 · What are a token and a context window? — LLM fundamentals and AI engineering
Further reading
- Anthropic: Effective context engineering for AI agents(opens in a new tab)29.09.2025. Choosing context, limits, and references instead of blindly growing the input.