“Make it better” is a difficult request for a person too. Does better mean shorter, more accurate, more persuasive or friendlier? AI fills those gaps its own way and may end up far from what you wanted.
Anthropic's article on context explains why selecting the information a model needs is part of designing the solution. You don't need to provide everything at once; you need material relevant to the task. Source: Effective context engineering.
Four things you can always specify
Explain the goal, provide the source material, set the boundaries and describe the output. For a customer email, that might look like this:
“Draft a reply to a customer asking about our training. The goal is to arrange a short conversation about their needs. Use the service description below. Don't add prices or dates that aren't there. Write in natural English, warmly and directly, in no more than 150 words. End with one simple question.”
That gives AI direction and leaves room for phrasing. An instruction's value doesn't depend on sounding technical.
Fix a specific part
If the first answer is too formal, identify the sentence that doesn't fit and show your own version. If it says little, ask it to tie each claim to a specific part of the service description. This also teaches you which inputs produce useful results.
Don't save a reusable instruction before testing it on several examples. One good email may be luck; different situations reveal whether the method holds up.
In AILithe's beginner training, we practise on your own work. The aim is to direct AI and judge its output yourself, without searching for a new “perfect prompt” every time.