Your first useful prompt

A useful prompt has four parts. Most failed prompts are missing at least two of them. Here's the structure and how to debug it.

The four-part structure

Every production-quality prompt has:

1. Role / frame. Who the model is answering as, and what kind of output is expected.
2. Context. The specific facts, examples, or documents the model needs to do the job.
3. Task. The concrete thing to do, in one or two sentences.
4. Format. The shape of the output — JSON schema, bullet list, specific sections.

Missing any one of these is where prompts go sideways.

A concrete example

Bad:

Summarize this article.

Useful:

You are an editor writing the lede paragraph for a tech newsletter read by
senior engineers. The readers skim. Lede paragraphs should be 2-3 sentences
that make a skeptical engineer want to read more.

Here is the article: <article>…</article>

Write one lede paragraph. Do not explain your reasoning. Do not say "This
article is about…". Start with a concrete claim or surprising fact.

The second prompt has frame (editor, skeptical engineers), context (the article), task (one lede paragraph), and format (2-3 sentences, start with a claim, no preamble). You'd get usable output.

Why prompts fail

When a prompt produces a mediocre answer, the diagnostic sequence is almost always:

1. Is the task crisp? Or did you give the model three things to do at once?
2. Is the context actually in the prompt? Or are you expecting it to "know" things about your domain?
3. Is the format specified? Vague format = lowest-common-denominator output.
4. Is the role coherent? Asking "an AI expert" for a plumbing tip gets you worse results than asking "a licensed plumber."

The prompt you'll use forever

Keep a personal template. Most engineers eventually settle on something like:

<role>You are…</role>
<context>…</context>
<task>…</task>
<format>…</format>
<examples>…</examples>  ← optional, see the few-shot lesson

The tags don't have to be XML. Headers work. Markdown works. The model doesn't care about the format of your structure. It cares that there is a structure.