How to write a prompt that actually works

Most weak prompts miss one of four things, rarely all four. Check them in order before you send anything non-trivial.

1. Who is the model being asked to be

Give a frame of reference and skip the personality. A senior backend engineer reviewing a pull request and a technical writer explaining this to a new hire give different answers to the same question, because each implies another reader and another level of explanation.

2. What exactly is being asked

Keep the task apart from the context. Here is a document, tell me about it is not a task. List the three riskiest assumptions in this document, with why each is risky is one. If you cannot state the task in one sentence with a verb in it, it is not specific enough yet.

3. What shape the answer should take

Length, format, structure: a list, a table, a paragraph, a fixed set of headings. Leave this out and the model chooses, usually a medium-length paragraph that hedges everything and aims at the average reader. That is rarely who is asking.

4. What to leave out

The instruction that saves the most rework is the negative one. Do not suggest a rewrite, only point out the issues, or skip anything that is a style preference and not a bug. It removes a whole category of unwanted output before it exists, so you do not edit it out afterwards.

Put it together

A prompt that covers all four seldom needs a fifth thing. Context and input go last, in a marked block. The prompt structure guide explains why order matters as much as content.