1. It completes; it doesn't comprehend your intent

A language model does one thing extraordinarily well: given everything written so far, it produces the most plausible continuation. When you prompt it, you are not filing a request with an assistant who knows you. You are setting up a situation and letting the model continue it.

This explains the most common disappointment: you knew what you meant, but the model only had what you wrote. If your words fit a hundred different intents, you get the average of those hundred intents, which reads as generic.

2. Everything it knows about you is in the conversation

Unless a platform has stored instructions or memory for you, the model starts every conversation knowing nothing about your job, your project, your standards, or your taste. It is not being difficult; the information genuinely is not there.

The practical rule: anything that would change the answer belongs in the prompt. Your role, your reader, your deadline reality, the format you need, if it matters, it must be written down. This is exactly the problem persistent profiles solve: writing the stable facts once instead of every time.

3. Words are instructions, examples are stronger instructions

Every part of your prompt steers the continuation: the words you choose, the tone you write in, even your formatting. Write sloppily and you have quietly asked for a casual register. Write a numbered list and you have hinted the answer should be structured.

This is why showing beats telling. One example of the output you want often outperforms a paragraph describing it, because the model continues patterns more reliably than it interprets descriptions.

4. Order and emphasis matter

Models pay attention to the whole prompt, but instructions land harder when they are explicit, near the task, and not buried in the middle of a long paragraph. A constraint whispered in passing ('oh and keep it short') competes with everything else you wrote.

Put the task up front, the constraints where they are unmissable, and repeat the one non-negotiable at the end if the prompt is long. Redundancy for the thing you care most about is not bad style. It is good engineering.