The Prompt Anatomy

Turn vague requests into prompts that work in one shot - a fill-in-the-blank, five-part brief.

Outcome: turn vague requests into prompts that get usable results in one shot — and know how to tell it worked.

Who it’s for: anyone using AI for real work who’s tired of “make this better” handing back mush.

The big idea: a good prompt isn’t a magic spell. It’s a brief — the kind you’d give a sharp new assistant. Five parts. Fill them in.

The five parts

1. Role — who you want the AI to be.

You are a [specific expert] who [relevant strength].

Sets the vocabulary, depth, and priorities. “A plain-spoken bookkeeper” answers differently than “a tax attorney.” Generic role → generic answer.

2. Context — what it needs to know (and nothing private it doesn’t).

Here’s the situation: [background]. It’s for [who]; it matters because [why]. Here’s the material: [paste / attach].

Most bad output is missing context, not a bad model. Tell it who the answer is for and why it matters — same facts, different reader, different answer. Give the right material, not all material — and leave out names, account numbers, anything you wouldn’t email.

3. Task — the one specific thing to do.

Do this: [single clear action].

One ask per prompt. “Summarize AND rewrite AND translate” gets you three mediocre things. Chain them in separate steps instead.

4. Format — what the output should look like.

Give it as [bullets / table / 3 short paragraphs / email], about [length], in a [tone] tone.

The model can’t read your mind about shape. Telling it up front saves the reformatting round-trip. Be concrete — “about 4 sentences” beats “short” — and say what you do want, not just what you don’t (“write in plain prose” works better than “no jargon”).

5. Example — show one good one (optional, but the biggest upgrade most people skip).

Good looks like: [sample]. — or — Match the style of this: [paste].

One example beats three paragraphs of description.

Fill-in-the-blank template

Role:    You are a ________________ who ________________.
Context: Situation: ________________.
         It's for ____________ ; it matters because ____________.
         Material: ________________.
         (Leave out: names, numbers, anything private.)
Task:    ________________   ← one thing.
Format:  ________________ , about ____ long, ____ tone.
Example: Good looks like: ________________.
If unsure: say so — don't guess.

Before → after

Before: “Write a follow-up email to a client.” → generic, wrong tone, you rewrite it anyway.

After:

Role: a warm, concise account manager. Context: 30-min call with a small-business client who’s nervous about timing. We agreed to push launch one week. Don’t mention budget. Task: draft the follow-up email. Format: 4 sentences max, friendly and reassuring, no jargon, end with one clear next step. Example: match the calm tone of “Thanks again for today — here’s where we landed…”

→ usable in one shot.

The five mistakes that cost you

No role → a generic average of the internet.
All context or no context → it loses the point, or it guesses.
Two tasks in one → split them.
No format → you spend prompting time on reformatting.
No example → the fastest fix you’re not using.

Did it work? (review standard)

Good output is usable with light edits, on-format, and about the thing you asked — not a lecture around it. If you’re rewriting from scratch, the prompt was missing Context or Example, not “the AI is bad.”

One move that cuts made-up answers

Add a line giving the AI permission to bail: “If you’re not sure, or it’s not in what I gave you, say so — don’t guess.” Giving it an out is one of the most effective ways to cut confident-but-wrong answers — it’s a fix the AI makers themselves recommend. (Then run the checklist anyway.)

Works on any model

The five parts work on every chatbot. One update for newer “thinking” models: you don’t need to tack on “think step by step” anymore — they reason on their own, so just ask clearly. (On older models, “think step by step” can still help.)

Safety note

Never paste private data (names, account numbers, health/legal details, anything under NDA) into Context. Describe it generically. What you paste can be stored.

Pair this with The “Is AI Wrong?” Checklist — a good prompt gets a good-looking answer, and good-looking is exactly when you need to check it.