The Anatomy of a Marketing Prompt: Audience, Goal, Format, Voice, Constraints

Let's talk about the single skill that decides whether AI is useful to your marketing or just a fancier way to generate stuff nobody wants to read. Prompting. Not the magic-incantation version you've seen on social, where someone swears their two-thousand-word mega-prompt changed their life. The actual job of it. What you're really doing when you type into the box.

So here's the whole episode in one sentence. A prompt is a brief. The same brief you'd hand a freelancer. The same brief you'd give a new junior hire on their first week. That's it. That's the mental model. Everything else we do today hangs off that one idea.

Anthropic, the company behind Claude, puts it really well in their own documentation. They say to think of the model as a brilliant but new employee who lacks context on your norms and your workflows. The more precisely you explain what you want, the better the result. Sit with that for a second. Brilliant, but new. The intelligence is there. The context is not. It doesn't know your audience. It doesn't know your brand. It doesn't know that your CEO hates exclamation points or that legal won't let you say the word guaranteed. You have to tell it. Just like you'd have to tell a sharp new hire who started Monday.

And the consequence of that is simple and a little brutal. Vague brief in, generic slop out. Specific brief in, on-brand draft out. The quality of what comes back is almost entirely a function of the quality of what you put in. When the output feels generic, it's not because the model is dumb. It's because your input was generic. We're going to come back to that idea over and over today, because it's the thing that unlocks all of this.

One quick note before we go deeper. Everything here is tool-agnostic. I'm going to say things like "the current ChatGPT" or "whatever Claude or Gemini you're using right now." I'm deliberately not pinning version numbers, because they change every few months and the principles don't. Paste a good brief into any of the big three and you'll get a better result. The job is the same everywhere.

This is also desk-level work. You, your hands, one chat window, one piece of copy at a time. We're not talking about automation, or agents, or hooking the model up to your whole database. That's later in the show. Today is the foundation: how to write one good brief for one asset. Get this right and everything fancy we build later actually works. Skip it and you're automating slop at scale, which is worse.

Okay. Let's build the anatomy. There are five parts to a good marketing brief, and I want you to define each one and see a real marketing example of each. Audience, goal, format, voice, and constraints. Five slots. Fill them in and you've gone from a lazy one-liner to something the model can actually run with.

Start with audience. Exactly who reads this thing. Their role, their knowledge level, what they care about, where they are in your funnel, and what objections they're carrying. Here's the trap. The model's default reader is everyone. And everyone is no one. When you don't say who it's for, the model writes for an imaginary average human, and average human copy is the blandest copy on earth. OpenAI's own best-practices guidance makes this point: specifying the role and the audience and the format gives you the most accurate, most relevant results. The audience line is doing a lot of that work.

So climb the specificity ladder. Don't say homeowners. Say millennial homeowners in suburban areas earning eighty to a hundred and twenty thousand dollars. Feel the difference? The first one could be anybody. The second one has a life, a budget, a set of worries. The model can write to that person.

And good news, marketers. You already own this artifact. It's your ICP, your ideal customer profile, which is just a one-line description of your perfect buyer. Or your buyer persona. You've probably got one in a deck somewhere. You don't have to invent the audience for every prompt. You paste it in.

Let me give you a real B2B example, B2B meaning business-to-business, you're selling to another company. Your audience line might read like this. A VP of Marketing at a two-hundred to five-hundred person B2B software company. Owns the pipeline. Reports to the chief revenue officer. Skeptical of, quote, another tool. Already got burned by software that promised automation and delivered busywork. Look how much that gives the model. Writing for a skeptic is a completely different job than writing for a believer. Writing for a beginner is different from writing for an expert. The audience line tells the model which person is on the other side of the screen, and everything downstream bends to fit them.

Second part. Goal. The job to be done. Not the topic, the outcome. There's a huge difference between "write about our new feature" and "get the reader to do this specific thing." A topic is a subject. A goal is an action. You want one clear outcome and the one action you want the reader to take, plus how you'll know it worked.

OpenAI's guidance again: outline what you want, who it's for, and why it matters. And Anthropic adds something subtle and useful, which is that giving the model the context and motivation behind your instructions helps it generalize. When it knows why, it makes better choices about the how. So don't just say "write a post." Say "the goal is to get the reader to book a twenty-minute demo. I'll know it worked if marketers save it and comment on it. The call to action should be specific, book a demo, not something soft like learn more."

And here's the rule that people fight me on. One goal per prompt. One. The moment you ask a single post to chase demos and newsletter signups and brand awareness all at once, you dilute all three. The model splits its attention, the copy gets mushy, and the reader doesn't know what you want from them, so they do nothing. Pick the one thing. If you need three outcomes, write three pieces. A brief with one clear job beats a brief with three competing ones every time.

Third part. Format. The shape. The container the words go in. Structure, length, channel limits, the output shape, whether that's a table or a list or sections or an email with a subject line or a two-hundred-and-eighty-character post. Here's a concrete fact about how these models behave. Without constraints, they ramble. You ask for "an email" and you get four hundred words when you needed a hundred and fifty. Not because it's wrong, but because you never told it where the edges were. So always state the exact length. Always.

Anthropic has the strongest formatting lesson I know, and it's counterintuitive, so let it land. Tell the model what to do, not what not to do. Instead of writing "do not use markdown," you write "write in smoothly flowing prose paragraphs." Same intent, but the positive version works far better, because you're pointing at the target instead of pointing away from a thousand things that aren't the target. We'll come back to this idea when we get to constraints, because it shows up there too. There's a related tip worth knowing: match your prompt style to the desired output. If you want tight punchy copy, write your prompt tight and punchy. The model picks up on that.

And learn your channel limits, because format is where they live. An email subject line wants to be around fifty characters. A Google ad headline is capped at thirty characters, which is brutally short. A meta description, the little blurb under your link in search results, runs about a hundred and fifty-five characters. A post on X is two hundred and eighty. Name those numbers in the prompt and the model writes inside the box instead of writing something you have to hack down later.

Here's a format line in practice. A LinkedIn post, one hundred and twenty to one hundred and fifty words. First line a hook, under ten words. No hashtags. Line breaks so it's skimmable. End with a one-sentence question as the call to action. That's a container. The model knows exactly what shape to pour the words into.

Fourth part. Voice. How it sounds. Your brand persona, your reading level, your point of view, we versus you, your tone. This is the bridge to the next episode, which is all about brand voice, so I'll keep it focused here. But I want to give you the one principle that matters most, and it's this. Voice is taught by showing, not telling.

Saying "be witty" is weak. The model has no idea what your witty is. Everybody's witty is different. But pasting three of your actual witty posts? That's strong. Now it has data. It can see your rhythm. There's even a nice move where you paste a few paragraphs of your best content and ask the model to analyze the sentence structure, the vocabulary, the rhythm, before it writes anything. You make it study your voice first.

If you did want to describe voice in words, here's what a good attempt sounds like. Confident, plain-spoken, like a senior marketer talking to a peer over coffee. Short sentences. Second person. Dry humor, no exclamation points. Reading level of a smart eighth grader. Never corporate, never hype-y. That's a decent voice line. But honestly? Three real examples beat that paragraph every time, and I'll prove it later in the episode.

One more thing on voice, and I'm going to mention it lightly because it really belongs to the next act of the show. There's a split worth knowing between your system prompt and your user prompt. Your brand identity and your voice, the stuff that's true for every single thing you write, can live somewhere saved. In ChatGPT that's your Custom Instructions or a Project. In Claude that's a Project's instructions. You set it once and it applies to every task. Then the specific job, write this one post, goes in the regular chat, the user prompt. Saved context is next act's territory, so I won't go deep. Just plant the flag: voice can be saved, so you're not retyping it forever.

Fifth part. Constraints. The guardrails. This is where you keep the model on the road. A few kinds. Hard limits, like word count, character count, number of variations. Must-includes, the things that absolutely have to appear, your target keyword, the offer, the product name spelled correctly, the call to action, a required disclaimer. Banned phrases, words you never want to see, like revolutionary, or game-changing, or in today's fast-paced world. And compliance, which matters enormously if you're in finance or health or legal. Don't say guaranteed returns. Include the disclaimer. Those aren't style preferences, those are how you stay out of trouble.

And one constraint that prevents a specific disaster. Factual grounding. You write something like, only use facts from the brief I gave you, and if you don't have a number, write the letters T-K in brackets instead of inventing one. T-K is an old editorial mark meaning "to come," a placeholder for a fact you'll fill in later. That one instruction stops the model from confidently making up a statistic, which it will absolutely do if you let it.

Now, there's a real tension in constraints, and you need to feel both sides. Constraints prevent slop. But over-constraining gives you robotic, lifeless copy. If you pile on forty rules, the model spends all its effort obeying rules and none of it writing well. The fix is the same lesson from format: pair your negatives with positive direction. Don't just ban the bad words, point at the good direction too.

Here's a constraints line that balances it. One hundred and fifty words max. Must include the phrase book a demo and the product name PipelineIQ. Do not use leverage, seamless, unlock, game-changer, or in today's fast-paced world. No em-dashes, no emoji. Only use the two numbers in the brief, don't invent figures. Tight, specific, and it blocks the exact stuff that makes copy smell like a machine wrote it.

So that's the five-part anatomy. Audience, goal, format, voice, constraints. Now I want to give you the highest-leverage move in all of prompting, because it deserves its own section. Examples. What the pros call few-shot prompting.

Let me define the ladder quickly. Zero-shot means you give no examples at all. The model guesses your style from the average of everything in its training, which is exactly why it comes back generic. One-shot means you give one example. Few-shot, sometimes called multishot, means you give several, usually two to five, so the model pattern-matches your style instead of guessing it. Anthropic's documentation says three to five examples gives the best results, and it gives you three properties to aim for. Make your examples relevant, meaning they mirror your real use case. Make them diverse, meaning they vary enough that the model doesn't latch onto some accidental pattern. And make them structured, meaning you label and separate them clearly so it knows where one ends and the next begins. Nice bonus: you can even ask the model to evaluate your examples for relevance and diversity, or to generate more in the same vein.

This is the show-don't-tell principle again, and it's worth repeating because it's the whole ballgame for voice. You cannot reliably describe your brand voice in words. Almost nobody can. But you can paste it. The examples do what the adjectives can't.

What makes a good marketing example? Your actual past work that performed. A real email that converted. A post that genuinely got engagement. Not aspirational copy, not the thing you wish you sounded like, the thing that actually worked. Grab two or three pieces that share a voice but cover different topics, so the model learns your style and not your subject. Plain text. Clearly labeled, something like, here are three of my past posts, match this voice. And a warning: don't paste three nearly identical posts, or the model over-fits one little quirk and repeats it forever. Three to five is the sweet spot. Even one or two strong examples beat zero by a mile. And past about five, you hit diminishing returns for this kind of desk work, so don't go overboard.

Alright. Let me hand you a set of practical techniques, all tool-agnostic, that make the anatomy sing.

First, role priming. Give the model a job title before you give it the task. You are a senior B2B software copywriter who's written for HubSpot and Notion. Anthropic notes that even a single sentence of role makes a real difference. But here's the honest caveat. Role sets tone. It does not grant expertise the model doesn't have. A job title won't rescue a thin brief. If you didn't give it facts, calling it a senior copywriter just gives you confident-sounding emptiness. Role is seasoning, not substance.

Second, delimiters and sections. Separate the parts of your brief so the model doesn't blur your instructions into your context into your examples. And for non-coders, good news: you don't need fancy tags or special syntax. Plain labels work just as well. Literally type the words audience, goal, format, voice, constraints, examples, each on its own line with your content under it. Plain headers. That's all the structure you need.

Third, give it source material. This is grounding. Paste in the product one-pager, the case study, the call transcript, the pricing page. Now the model writes from your facts instead of inventing its own. Give it background about your business and your audience early, and it pays off in every later prompt in that chat. One specific tip: when you're pasting a long document, put the long material at the top, above your actual question. With big pasted documents that ordering can meaningfully improve the quality of what comes back.

Fourth, ask for multiple variations. Give me three versions of the hook, each from a different angle. One prompt, and suddenly you've got an A/B menu to choose from instead of a single take-it-or-leave-it draft.

Fifth, and this is a mindset shift more than a technique. Iterative refinement. The prompt is a conversation, not a one-shot. You don't write the perfect brief and walk away. You start, you read what comes back, and you refine. Tighten paragraph two. Make the call to action harder. Cut thirty words. More like example two. You're directing, not restarting. Every one of those follow-ups is you steering, and steering is faster than rewriting from scratch.

Sixth, chain your outline to your draft. Ask for the outline or the angles first. Approve it, edit it, then ask for the full draft. You catch a wrong direction while it's cheap to fix, before the model's written eight hundred words down the wrong path. There's a sharper version of this, a self-correction loop. Draft this email. Now critique it as a skeptical chief marketing officer. Now rewrite it, fixing those critiques. The model plays both writer and tough editor. Just keep it framed as hand-driven turns in one chat. You're typing each step. This is not automation, it's you running a conversation with structure.

Let me give you the headline rules straight from the official documentation, because they're a tidy summary of everything so far. Be specific. Give examples. Tell it what to do, not just what to avoid. Use structure and delimiters. Add context and motivation. That's the spine of both Anthropic's and OpenAI's guidance, distilled.

And here's the golden rule, a pull-quote from Anthropic that I want you to actually remember. Show your prompt to a colleague with minimal context and ask them to follow it. If they'd be confused, the model will be too. That's your test. If a smart human who doesn't live in your head couldn't execute your brief, neither can the machine. Confusion in, confusion out.

Okay. Theory's great. Let me show you the whole thing in motion. We're going to build one LinkedIn post, before and after.

Here's the before. The lazy one-liner. Write a LinkedIn post about our new analytics feature. That's it. And here's what comes back, more or less every time. It opens with in today's fast-paced digital landscape. It calls the feature game-changing and seamless. Three emoji sprinkled in. Some vague benefit bullets that could apply to any product on earth. And it signs off with let's connect, followed by five hashtags. It sounds like every company. You could paste your competitor's name over yours and it would fit perfectly. That's the tell, and we'll name why in a minute.

Now the after. Same feature, but we load the five-part anatomy. Role: you are a senior B2B software copywriter. Audience: in-house demand-generation marketers at two-hundred to five-hundred person software companies, skeptical, drowning in dashboards they don't trust. Goal: get them to book a twenty-minute PipelineIQ demo, and I'll know it worked by saves and comments. Format: one hundred and twenty to one hundred and fifty words, a hook under ten words, one idea per line, a question as the call to action, no hashtags. Voice: confident, plain, peer-to-peer, short sentences, second person, dry humor, no exclamation points, eighth-grade reading level. Constraints: must include the phrase book a demo and the name PipelineIQ, ban leverage and seamless and unlock and game-changer and in today's fast-paced world, no em-dashes, no emoji, don't invent stats. And examples: paste in three of your past posts that performed.

Now watch what each element actually changes, because this is the part that makes it click. The audience line swaps those generic benefits for the skeptical, dashboard-fatigued angle, copy that speaks to a real worry. The goal line gives you a real call to action instead of let's connect. The format line kills the wall of text and hands you a hook plus skimmable lines. The voice line drops the corporate register and makes it sound like a person. The constraints strip the slop words and block any fabricated numbers. And the examples? The examples are the single biggest jump in on-brand-ness of anything on that list. That's the move that makes it sound like you.

Then you keep going, because it's a conversation. Give me two more hook options. Make the call to action harder. You're refining, not restarting. Two minutes of follow-up and you've got something you'd actually publish.

A couple of variations on that worked example. If you were writing an email instead of a post, you'd use a copywriting framework called PAS, problem, agitation, solution, inside your goal and format. More on PAS in a second. And if you were writing a landing page hero, the headline up top of a page, the classic pitfall is describing features instead of benefits. Frameworks like PAS help because they structure the copy around what the reader actually cares about, not around what your product technically does.

Now let's talk about the one pitfall I want you to anchor everything to. Slop.

Slop was the 2025 Word of the Year. Both Merriam-Webster and the American Dialect Society named it. AI slop is content that reads as lacking effort, quality, or meaning. And here's the thing that should make you feel better, not worse: even the model makers admit it in their own documentation. They say that without guidance, models default to generic patterns. They converge toward generic, on-distribution outputs. That's the technical phrase for the AI-slop aesthetic, coming from the people who built the models.

Here's the why underneath it, and it's the most important sentence in this episode. With no constraints, the model returns the statistical middle of everything it has ever seen. The average. The center of the blob. And the average of all marketing copy is, by definition, the most generic marketing copy. Specificity is the force that pulls the output off that average. That's all your brief is really doing. It's dragging the model away from the middle toward your specific, real, particular thing.

Let me give you a listen-for checklist, because once you can hear slop, you can't unhear it. Opener cliches: in today's fast-paced digital world, picture this, as a business owner you know. Buzzword bingo: delve, leverage, foster, empower, unlock, seamless, robust, cutting-edge, game-changer, transformative, revolutionary. Spatial metaphors where there's no actual space: tapestry, landscape, ecosystem, realm, beacon, journey. The "it's not just X, it's Y" construction. The "no this, no that, just this" rhythm. Hedging language: it is important to consider, generally speaking. Listicle-brain, those rigid bold-header-colon-description bullets, intro then point then point then point then conclusion. Monotone rhythm, every sentence the same fifteen to twenty words, no fragments, no variation. Transition tics: furthermore, moreover, additionally, in conclusion, at the end of the day. Vague fake experience with zero specifics. And em-dash overuse.

But here's the master diagnostic, the one test that catches all of it. No specifics. Slop never names a real number, a real customer, a real moment. So run this check. If you can swap your company name for a competitor's and the copy still fits, it's slop. Every time. If it could be anyone's, it's no one's.

And now you see why the fix is upstream. The output is generic because the input was generic. The five-part anatomy is the de-slopper. Audience plus voice plus examples plus banned-phrase constraints, those four together pull the model off the average and onto your brand. And the workflow that holds it all: human strategy, then AI draft, then human edit. Never raw publishing. You strategize, the model drafts, you edit. The machine sits in the middle, never at the end.

One quick flag while we're here, and I'll keep it short because it gets a full episode later. Hallucination. Ask the model for facts or stats or quotes with no source material, and it may confidently invent them. Whole-cloth, sounds-totally-real, completely made up. The fix for now is grounding. Paste the source, or tell it to write T-K in brackets for any number it doesn't actually have, and fact-check before you publish. That's enough for today. Just know it's a thing.

Let me address the elephant in the room: all those prompting frameworks with the catchy acronyms. You've seen them. RTF, RACE, CRISPE, RODES, TRIM. Here's my honest take. They're all repackagings of the same handful of ideas, role plus task plus context plus format plus examples. Pick one. Don't collect them.

But quickly, so you know the menu. RTF is role, task, format. Simplest, fastest, a good beginner on-ramp, but it's missing context and examples, which are the two things that fix generic output. RACE is role, action, context, expectation. It adds the context slot, and context is very often the missing piece behind a generic answer, so RACE is a solid default. CRISPE is heavier, and honestly, blogs disagree on what the letters even stand for, which tells you something about how loosely standardized these things are. The useful bits in it are Personality, which is voice, and Experiment, which is asking for variations. RODES is role, objective, details, examples, sense-check. I like this one because it bakes in examples, the few-shot move, and a sense-check, a self-review step. It's the closest to our five-part anatomy plus iteration. And TRIM is task, relevant context, intent, measurable criteria. It's marketing-specific and strong on measurable success criteria, which ties right back to our goal slot.

The honest framing: acronyms are scaffolding, not magic. Internalize the anatomy, then use whichever mnemonic helps you remember it. The map is not the territory.

Now, separate but adjacent, there's a different family of frameworks you should not confuse with the prompting ones. Copywriting frameworks. These are about the structure of the copy itself, and they feed into your format and goal slots. AIDA: attention, interest, desire, action. PAS: problem, agitation, solution, which has strong momentum because it cuts straight to the emotional core, fast. And BAB: before, after, bridge. Pro move, you can hybridize them. Use PAS for the hook to grab the emotion, then AIDA for the close to drive the action. And the best part, you can name the framework right inside your prompt. Just write, structure this email using PAS. The model knows these. It'll comply.

Before we land the plane, let me run through the common mistakes, each with its fix, because knowing the failure modes is half the skill.

One. Too short and vague, treating the AI like a search engine. You type three words and expect magic. The fix is the whole five-part anatomy. Give it a brief.

Two. Conflicting instructions. Asking for punchy and comprehensive in the same breath. Short, but cover everything. The model can't do both, so it splits the difference and gives you mush. Pick a lane.

Three. Asking too much at once. The overstuffed prompt that wants a strategy and a draft and ten variations all in one shot. Break it into smaller prompts. Outline first, then draft.

Four. Not giving examples. We covered this. Few-shot, three to five. It's the biggest lever you've got and most people skip it.

Five. Not iterating. Accepting the very first draft like it's final. Treat it as a conversation. The first draft is a starting point, never the finish line.

Six. Copy-pasting someone else's mega-prompt without adapting it. That two-thousand-word prompt you found on Reddit carries a stranger's audience, a stranger's voice, a stranger's constraints. It was tuned for their brand and their funnel, not yours. The fix: strip it down to the anatomy. Swap in your audience, your voice, your examples. Delete what doesn't apply. And then test it on your brand and your funnel, because that's the only test that matters. A prompt that crushed it for someone else's product means nothing until it works for yours.

Seven. Over-constraining. We touched it earlier. Pile on too many rules and you get robotic copy. Lead with what to do, keep only the constraints that genuinely matter, and let the model breathe.

Here's the meta-point I want to close on, and it's the mindset that separates dabbling from actually directing AI. A prompt is a reusable asset. Not a one-off. The whole shift happening across 2025 and into 2026 is from one-off experimentation to saved playbooks. You wrote a great brief that produced a great LinkedIn post? Don't throw it away. Save it. Reuse the skeleton for your next post, your next channel, your next campaign.

So let me leave you with that skeleton, spoken plainly so you can build it yourself. You're going to set up slots, top to bottom. A role line, where you put seniority plus specialty. An audience line, who reads it, and you paste your ideal customer profile right there. A goal line, one action plus how you'll know it worked. A format line, the asset type, the exact length, the structure, the channel limits. A voice line, your tone, your point of view, your reading level, and remember, better to show it in the examples than describe it. A constraints line, your hard limits, your must-includes, your banned phrases, and that grounding instruction, don't invent stats, use the facts below or write T-K. Then a context block, where you paste your source material, the one-pager or the case study or the pricing page. Then an examples block, three of your best past pieces. And finally, the kicker line: now write it, then give me two alternative hooks.

That's the skeleton. Pair it with the iteration habit, save the filled-in version per channel, and you've got a system. But hear me on the last thing, because it's the thing people forget. The template is inert until you load it with your specifics. An empty skeleton is worth nothing. It's the audience, the voice, the examples, your real ones, that make it work. Test it on your brand and your funnel. Always.

That's the anatomy of a marketing prompt. Audience, goal, format, voice, constraints, plus examples as your secret weapon and iteration as your habit. A prompt is a brief. Write it like one, and the slop never shows up in the first place. Next time, we go deep on voice, on capturing what makes your brand sound like your brand and nobody else's. See you then.