The Four-Part Prompt Structure (With Before and After)

Four parts. Skip one and you get a specific, predictable failure — which is useful, because it means you can diagnose bad output by working out which part was missing.

Role, context, task, format. Here's what each does, what breaks without it, and five real prompts rewritten.

The short version

Part	What it does	What breaks without it
Role	Sets the perspective and register	Output pitched at the wrong level
Context	Supplies what the model can't know	Generic output — the most common failure
Task	States what you want done	It answers a related question instead
Format	Specifies what comes back	Wrong length, wrong shape, one option when you wanted twenty

Context is doing most of the work. Role is doing the least, despite getting the most attention.

Part 1 — Role

The "act as an expert copywriter" move. Genuinely useful, and less important than its reputation suggests.

What it does: narrows the register and the assumed vocabulary. Asking for an explanation "as a technical documentation writer" versus "as a friendly explainer" produces meaningfully different output.

What it doesn't do: make the model more knowledgeable. Telling it to act as a world-class expert doesn't add expertise it doesn't have — it just changes how confidently it phrases things, which is occasionally the opposite of helpful.

When it earns its place: ambiguous tasks where register matters. "Review this" is ambiguous; "review this as a sceptical editor looking for unsupported claims" isn't.

When to skip it: clear, mechanical tasks. Extracting phrases from a transcript doesn't need a persona.

One caution: "act as a doctor/lawyer/financial advisor" produces confident output in that register without producing actual expertise. The confidence is the risk.

Part 2 — Context

The part that matters most and gets skipped most.

What belongs here:

• Who the output is for, specifically — their situation, not their job title
• What you've already tried or ruled out
• Constraints: length, tone, things you can't say, brand rules
• What you'd consider a failure
• Any relevant background the model can't infer

Why it dominates: every gap in your context gets filled with the most typical assumption. Write for "marketers" and you get output for an average marketer who doesn't exist. Write for "a solo consultant whose client's emails started landing in spam last week, technically confident but has never edited a DNS record" and everything downstream changes — the assumed knowledge, the tone, the urgency, how much each concept gets explained.

The test: could someone else read your context and produce roughly the right thing? If not, the model can't either.

Part 3 — Task

What you actually want done, stated as an instruction rather than a topic.

Weak: "help me with my email subject lines"
Strong: "generate twenty subject line options across four named frameworks"

The distinction: a topic invites the model to decide what would be helpful. An instruction tells it. Given a topic, it usually produces an overview — which is rarely what you wanted.

Be explicit about the verb. Generate, extract, rewrite, critique, compare, summarise, translate. Each produces a different kind of output, and "help me with" produces whichever it guesses.

One task per prompt where possible. Asking for research and a draft and a headline in one go produces three mediocre things. Chain them instead.

Part 4 — Format

What comes back, and how much of it.

Specify: length, structure, how many options, whether it's prose or a list, and any structural rules.

Why it matters more than it sounds: without a format, you get the model's default, which is an essay of moderate length with a summary at the end. That's rarely the shape you needed.

Useful format instructions:

• "Twenty options as a plain numbered list, no commentary"
• "Under 400 words, three short paragraphs, no headings"
• "A table with these columns: [x], [y], [z]"
• "Vary the section lengths — not every section the same size"

That last one is worth building into any drafting prompt. Uniform sectioning is the most visible tell of unedited AI output, and the model won't vary it unless told. The other tells.

Five rewrites

1 — Blog outline

Before: "Write an outline for a blog post about email deliverability."

After: "You're an SEO content strategist. I'm writing for marketing managers at small SaaS companies who've set up SPF and DKIM but still land in spam — frustrated, and suspicious they've been sold a myth. My position: authentication proves identity, not desirability, and the two factors that actually decide placement are list quality and sending behaviour. Produce an H2 outline of six to eight sections with a rough word count for each. Vary the lengths — some sections should be 200 words, some 600. No section that summarises the introduction."

2 — Customer research

Before: "What language do customers use about email deliverability?"

After: "Here are 50 support tickets and reviews from my customers. [pasted] Extract the exact phrases they use to describe the problem, the words they use for the outcome they want, and the objections that recur. Quote their language directly rather than paraphrasing. Use only this material — don't add phrases from general knowledge."

3 — Rewriting for a different audience

Before: "Make this simpler."

After: "Rewrite this section for a non-technical business owner who knows what email marketing is but has never seen a DNS record. Keep every piece of information — don't cut anything, just change the assumed knowledge. Same length. Use one concrete analogy for the technical concept."

4 — Critique

Before: "What do you think of this?"

After: "Review this draft as a sceptical editor. Identify: any claim stated without support, anywhere the argument hedges instead of committing, and any section that could be cut without loss. Don't rewrite it — list the issues with the specific line for each."

5 — Variants

Before: "Write me a CTA."

After: "Give me eight CTA phrasings for this email. [pasted] Each should state what the reader gets rather than what they do — 'see the three new colours' not 'shop now'. Include a specific detail in each: a number, a time cost, or a name. Plain list, no commentary."

The pattern across all five: the "after" versions aren't cleverer. They're specified.

One message or several?

Several, for anything with multiple stages.

Why: each response gives you a checkpoint. If the outline is wrong, you fix it before 2,000 words get built on it. Cramming research, outline, and draft into one prompt means correcting all three at once.

The sequence that works: context and material first, then task one, review, then task two.

When one message is fine: single-stage tasks. Extraction, rewriting, generating variants.

How specific should you be?

Specific about what matters, silent about what doesn't.

Over-specifying — dictating every sentence — means you've written it yourself with extra steps, and you've removed the model's ability to surprise you usefully.

Under-specifying gets you the generic default.

The line: specify anything you'd reject the output for. If you'd send it back for being too long, specify length. If you'd send it back for hedging, say don't hedge. If you genuinely don't mind, leave it open.

And when you don't know what you want, that's worth noticing. The difficulty of writing the prompt is often the signal that you haven't decided yet — which no amount of prompting fixes.

Frequently asked questions

Does telling AI to "act as" help?
Modestly, and less than its reputation suggests. Role prompting narrows the register and assumed vocabulary, which genuinely helps on ambiguous tasks where tone matters — "review this as a sceptical editor" is more useful than "review this." What it doesn't do is add expertise the model lacks; telling it to act as a world-class expert changes how confidently it phrases things without changing what it knows, and that confidence can be actively unhelpful.

How specific should a prompt be?
Specific about anything you'd reject the output for, and silent about the rest. If you'd send it back for being too long, specify length. If you'd send it back for hedging, say so. Over-specifying means you've written it yourself with extra steps and removed any chance of a useful surprise. Under-specifying gets you the generic default, which is what the model produces when you leave a decision to it.

Should I prompt in one message or several?
Several, for anything with stages. Each response is a checkpoint — if the outline is wrong you fix it before two thousand words get built on it, whereas cramming research, outline, and draft into one prompt means correcting all three at once. Single-stage tasks like extraction, rewriting, or generating variants are fine in one message.

What to do next

Take a prompt you used this week and check which of the four parts it's missing.

For most people it's context — the prompt states a task and a format and assumes the model knows the situation.

Add two specific sentences about who the output is for and what they already know. That's the highest-return thirty seconds available in prompting.

Free: The marketing prompt library.

---

Related guides

• Prompting for marketers — the wider picture
• Few-shot prompting — adding examples to the structure
• The content brief — the same logic at article scale
• Building a prompt library — reusing what works