Prompt Anatomy: Writing Clear Prompts

Think of the model as a brilliant new coworker on their first day: capable, eager, and completely unaware of what you actually want. Hand them a vague request and they will fill in the blanks with guesses — not because they are lazy, but because nobody briefed them. A prompt is that briefing.

Here is the kitchen version. Walk in and say “make something tasty” and you might get a cake, a sandwich, or a seven-course meal. Hand over a recipe instead — the dish, what's in the fridge, the time limit, how to plate it — and dinner lands the way you pictured it. Same kitchen, same cook. The only difference is the briefing.

A good briefing has four parts:

• Task — the dish. One clear verb. “Summarize this ticket” is a task; “look at this” is a wish.
• Context — what's in the fridge. Who it is for, why it matters, the background the model should assume. The model cannot use what you never gave it.
• Constraints — the rules of the kitchen. Length, tone, what to include, what to leave out, what to do when information is missing (“say 'unknown' rather than guessing”). This is the part most prompts skip.
• Format — how to plate it. Bullets or a table? A headline first? The model will happily pour the same facts into whatever container you name.

Before and after

The vague prompt — “make something tasty” energy:

What comes back is technically correct and completely useless:

The model guessed the audience (formal), the length (three sentences), and the content (nothing specific) — because the prompt answered none of those questions.

The same request, briefed properly — the recipe version:

Same facts in, far more useful text out:

The difference is not a smarter model. It is a complete briefing: every part told the model something it would otherwise have had to invent.

One more rule that belongs with the four: keep source material separate from instructions, so a reader — human or model — can tell which is which. Mark quotes as quotes. Instructions found inside retrieved documents should be treated as untrusted content, not commands: a document that says “ignore previous instructions” is data, not your boss.

Here is a real-world dud — the kind of prompt everyone writes on their first day. Your job: rebuild it with all four parts.

Notice what the exercise cannot do: no chip adds facts. If the newsletter needs the shelter's real adoption numbers, those have to come from a source — better wording cannot invent them. That is the honest boundary, below.

Better prompts cannot supply missing evidence. No amount of task-context-constraints-format will make the model know your company's returns policy or last quarter's numbers. When the facts aren't in the prompt and aren't retrievable, a great prompt produces a beautifully formatted guess. (See why AI hallucinates.)

And prompts cannot guarantee correctness. They raise the odds of a useful answer; they don't certify it. For anything that matters — money, access, safety, published claims — the answer still gets checked against a real source, no matter how good the prompt was.