When an answer is too generic, a longer prompt is not automatically the solution. Often it lacks information you would give a colleague: the purpose, established facts and remaining uncertainty.
Four things to include
- Goal: what should you be able to do with the answer?
- Context: who is it for and what information matters?
- Boundaries: what must not be assumed or added?
- Format: a short email, a table or a list of open questions?
Anthropic’s prompting documentation emphasises clear instructions, context and examples. This is a useful starting point across tools, although you still need to judge the result against your own task.
An example customer email
“Write a short email to a customer about a delayed delivery. The original date was 12 October. A new date has not been confirmed. We expect an update from the supplier on Friday. Explain what we know, do not promise a delivery date and say we will update them on Friday. Be direct and friendly, in no more than 100 words.”
This provides enough direction for a draft and explicitly states what is unknown. It will not prevent every error, but it makes the checking criteria clear.
Improve the result specifically
Name what needs changing. “Less formal” is more useful than “make it better”. More specific still: “Start with the delay, remove the long introduction and do not change the facts.”
For recurring work, keep a good example. Only add data that your organisation allows you to use in that tool.
I would start with a few frequent tasks and refine the instructions based on the results. That creates something useful for your actual work.
Sources and further reading
Explanation and commentary based on the sources below. Examples are illustrative.