The Anatomy of a Great Prompt: 6 Parts That Change Your Results

Ask ten people for a “prompt formula” and you’ll get ten acronyms. They mostly point at the same idea from different angles. Rather than memorize an acronym, it’s more useful to understand the parts a prompt can contain, why each one exists, and when to bother including it.

This guide breaks a prompt into six building blocks. Think of them like ingredients, not a recipe you must follow every time. A quick question needs one or two. A high-stakes deliverable might use all six. By the end you’ll have a reusable template and, more importantly, the judgment to know which blocks a given task actually needs.

The Six Building Blocks at a Glance

Here’s the whole model in one view. Read top to bottom — each block adds precision, and you can stop whenever the request is clear enough.

  ┌─────────────────────────────────────────────────────────┐
  │ 1. ROLE        Who should the AI be?                     │
  │    "Act as an experienced hiring manager."              │
  ├─────────────────────────────────────────────────────────┤
  │ 2. TASK        What one thing should it do?             │
  │    "Review this résumé bullet."                         │
  ├─────────────────────────────────────────────────────────┤
  │ 3. CONTEXT     What background does it need?            │
  │    "For a senior data-analyst role in fintech."        │
  ├─────────────────────────────────────────────────────────┤
  │ 4. FORMAT      What should the output look like?        │
  │    "Give a rewrite, then 2 lines on why."              │
  ├─────────────────────────────────────────────────────────┤
  │ 5. EXAMPLES    What does 'good' look like?             │
  │    "Match the style of this sample: ..."               │
  ├─────────────────────────────────────────────────────────┤
  │ 6. CONSTRAINTS Rules, limits, things to avoid.          │
  │    "Under 20 words. No buzzwords like 'synergy.'"      │
  └─────────────────────────────────────────────────────────┘

Let’s take each one in turn, with the reasoning behind it — because knowing why a block helps is what lets you improvise later.

Block 1: Role — Who the AI Should Be

A role tells the model which “voice” and knowledge frame to draw on. “Explain interest rates” and “As a patient financial educator explaining to a teenager, explain interest rates” pull from different registers. The second is simpler, warmer, and better scaffolded.

Roles work because the model has absorbed countless examples of how a teacher, a lawyer, a copy editor, or a skeptical reviewer tends to write. Naming a role activates those patterns.

When to use it: When tone, depth, or perspective matters. A hiring manager reviews a résumé differently than a friend would.

When to skip it: For plain factual questions. “As a geographer, what’s the capital of Peru?” adds nothing.

A common mistake is stacking grand titles — “world-renowned Nobel-winning expert.” Impressive adjectives don’t add competence; they mostly add flowery language. A plain, specific role (“a copy editor who writes for a general audience”) beats a decorated one.

Block 2: Task — The One Clear Verb

The task is the heart of the prompt: the single action you want. The key word is single. Prompts fail most often not because the task is complex but because it’s actually three tasks crammed together and the model does one well and two poorly.

Strong tasks start with a concrete verb: summarize, rewrite, compare, list, draft, translate, critique, plan, classify. Vague tasks start with mush: help with, do something about, look at.

Compare:

Weak:   Help me with my LinkedIn.
Strong: Rewrite my LinkedIn headline to emphasize data analysis
        and make it fit in 120 characters.

If you genuinely need several things, either break them into a numbered list within one prompt or — better — do them in sequence across a conversation, reacting to each result before moving on.

Block 3: Context — What It Can’t Know Unless You Say

Context is everything about your situation the AI can’t see: who the reader is, what happened before, what you’ve already tried, what matters and what doesn’t. This is the block people skip most and regret most.

The model has no access to your world. It doesn’t know your customer is angry, that your last three emails went unanswered, or that your boss hates exclamation points. Supplying that context is often what turns a generic answer into a tailored one.

A practical way to generate context: imagine the AI is a competent temp who just walked in. What would you tell them so they don’t embarrass you? That’s your context.

Context example:
"This reply goes to a long-time client who's usually easygoing
but is frustrated we missed a deadline. I want to keep the
relationship warm while being honest that we slipped."

Notice how much that single paragraph constrains the output in a good way. Tone, stakes, and history are all encoded.

Block 4: Format — Shape the Output Before You See It

Format tells the AI what the answer should look like: a table, bullet points, a numbered plan, a single paragraph, an email, a word count. Without it, the model defaults to its favorite habit — often long, hedged prose you then have to trim.

Specifying format saves the most editing time of any block. A few high-leverage format instructions:

• “Answer in a table with columns X, Y, Z.”
• “Give me exactly 5 options, one line each.”
• “Keep it under 100 words.”
• “Start with the conclusion, then three supporting bullets.”
• “No preamble — just the rewritten text.”

That last one is underrated. Models love to open with “Certainly! Here’s a revised version:” — telling it to skip the throat-clearing gives you paste-ready output.

Block 5: Examples — Show, Don’t Just Tell

Sometimes the fastest way to explain what you want is to show one. If you paste a sample and say “match this style,” the model has a concrete target instead of guessing what “professional but friendly” means to you specifically.

This is powerful for matching voice. Give it two of your past emails and it can draft a third that sounds like you. Give it one product description you like and it can write ten more in that mold.

Here are two headlines in our brand voice:
- "Coffee that respects your morning."
- "Small-batch. Big opinions."
Write 5 more headlines for a new cold-brew line in this exact voice.

Even a single example dramatically narrows the range of outputs. It’s the difference between describing a color and pointing at it.

When to skip it: When the task is generic and you don’t have a strong preference. Don’t manufacture examples you don’t actually care about.

Block 6: Constraints — The Guardrails

Constraints are the rules and limits: length caps, words to avoid, a budget, a reading level, required elements, forbidden ones. They keep the output inside the lines.

The best constraints are checkable. “Make it professional” is a wish. “No exclamation points, under 80 words, and don’t mention pricing” is a spec you can verify at a glance.

Useful constraint types:

Constraints also protect against the model’s default tendencies — over-explaining, hedging, or drifting off topic. A single “be concise; skip the disclaimers” can transform an answer.

Putting It Together: A Reusable Template

Here’s a fill-in-the-blanks template using all six blocks. Copy it, delete the lines you don’t need, and keep it somewhere handy.

Role:        Act as a [role] who [relevant trait].
Task:        [One verb] the following: [the thing].
Context:     This is for [audience/situation]. Background: [key facts].
Format:      Respond as [format], [length limit].
Examples:    Match this style/sample: [paste].
Constraints: Must [required]. Avoid [forbidden].

And here it is filled in for a real task — writing a product update email:

Role:        Act as a product marketer who writes clear, no-hype updates.
Task:        Draft an announcement email for a new "dark mode" feature.
Context:     Audience is existing users, mostly non-technical, who asked
             for this feature repeatedly. We want them to feel heard.
Format:      Short email: subject line + 3 short paragraphs + one button
             label. Under 140 words total.
Examples:    Our voice is warm and plain, like: "We heard you. It's here."
Constraints: No buzzwords, no exclamation points, mention it's free.

The output from a prompt like this needs almost no editing, because you’ve made nearly every decision in advance instead of leaving them to chance.

Before and After: Watching the Blocks Work

Let’s see the cumulative effect. Same goal — a caption for a small bakery’s social post — built up one block at a time.

Task only:

“Write a caption for a bakery.” → Generic, could be any bakery on earth.

+ Context:

“Write a caption for a small family bakery known for sourdough, posting a photo of a fresh loaf.” → Now it’s about this bakery.

+ Tone/Role:

“…in a warm, homey voice, like a baker talking to neighbors.” → Now it has personality.

+ Format + Constraints:

“…under 20 words, one line, no more than one emoji, include the word ‘sourdough.’” → Now it’s postable as-is.

Each block removed a category of wrongness. That’s the mental model to carry: every block you add closes off a way the answer could miss.

How Many Blocks Do You Actually Need?

Here’s the honest answer most guides won’t give you: usually two or three. For everyday use, Task + Context + Format carries most of the weight. Role, examples, and constraints are situational upgrades you reach for when the first three aren’t enough.

A simple decision rule:

• Casual question? Just the task.
• Something you’ll actually use? Task + Context + Format.
• High stakes, or the first attempts missed? Add Role, Examples, and Constraints as needed.

Don’t pad prompts to feel thorough. A bloated prompt full of unnecessary blocks can actually muddy the request. Precision beats length. The goal isn’t to use all six — it’s to include exactly the ones that remove real ambiguity.

The Opposite Failure: The Over-Stuffed Prompt

Once people learn about the six blocks, a predictable overcorrection follows: they cram all six into every request, pile on adjectives, and end up with a bloated paragraph that actually performs worse than a lean one. It’s worth understanding why, because it saves you from trading one bad habit for another.

A prompt is a set of instructions competing for the model’s attention. When you add a block that carries no real information — a decorative role, a constraint you don’t care about, an example that doesn’t match your goal — you don’t add precision. You add noise. The model now has to weigh a fake requirement against your real ones, and sometimes it guesses wrong.

Here’s an over-stuffed prompt for a simple task:

Act as a world-class, award-winning, senior communications strategist
with 20 years of Fortune 500 experience. Leveraging your deep expertise
and thought leadership, craft a truly compelling, best-in-class, highly
engaging yet professional message that resonates deeply to inform my
neighbor that a package of theirs was delivered to my door by mistake.

All of that scaffolding is fighting to justify itself, and the result is usually overwrought — a grand corporate memo about a misdelivered parcel. The lean version wins easily:

Write a short, friendly note to my neighbor letting them know a package
of theirs was delivered to my door by mistake and they can pick it up
anytime. Casual and warm, 2-3 sentences.

Task, tone, and constraints — three blocks, no filler — and it’s exactly right.

The lesson: every block should earn its place by removing a real ambiguity. Before you add one, ask “if I delete this, does the answer get worse?” If not, cut it. Precision comes from including the blocks that matter and excluding the ones that don’t. A prompt is a scalpel, not a snowball — you’re aiming for the smallest set of instructions that fully specifies what you want.

Key Takeaways

• A prompt has six possible parts: role, task, context, format, examples, constraints. Each closes off a specific way the answer could go wrong.
• Task and context do the heavy lifting. Start there; add the rest only when needed.
• Constraints and formats should be checkable, not wishes — “under 80 words” beats “make it short.”
• One example often beats a paragraph of description. Show the model your target when style matters.
• More blocks isn’t better; the right blocks are. Match the effort to the stakes.

Once you see prompts as assembled from parts rather than typed as one shot, editing them gets easy. A bad answer points straight at the missing block — add context, tighten the constraint, show an example — and you’re done.