The Six Parts of a Working Prompt

The six fields of a real brief

The core claim: when an AI assistant returns generic, "slop" copy, it's usually not failing — it's filling in blanks you left. A prompt is a brief, and a brief has six fields: audience, goal, format, voice, constraints, and examples. Most people fill in one.

Using a small, real example — an email telling existing clinic customers about a new text-reminder feature — each field is worked through in turn. Audience means naming one real person (a practice manager, not "our customers"), knowing where she stands with you already, and knowing what she already believes, so the email argues the right thing instead of restating what she knows. Goal means one action moved, not several stacked asks that cancel each other out. Format means specifying length, structure, and where it's actually read — in this case, a phone screen, which forces the whole argument into the first line. Voice means describing your writing by behavior ("we say what it does before why it matters") or, better, pasting a real sample and asking the tool to match it, rather than reaching for adjectives like "friendly and professional" that describe nothing. Constraints are the fence: what must appear, what must never be claimed (like an invented no-show percentage), and which words are banned. Examples are the strongest lever of all — pasting real prior work of the same type and asking the tool to follow that pattern, because a pattern shown beats a pattern described.

The assembly method: one file, task-specific fields on top, stable fields (voice, constraints, examples) underneath, filled out in a single pass and judged against the brief itself rather than against taste. The core discipline covered at length: when a draft fails, fix the field that's thin, not the output — because a correction typed into a chat dies when the tab closes, while a fixed field survives into every future run. The chapter also names the trap of arguing with a draft turn by turn until it degrades, and gives the tell for catching it: the moment you're giving the same note twice.

What shipped this month

Ad platforms are taking bidding decisions off marketers' desks — ChatGPT's ad platform switched on <a href="https://www.jumpfly.com/blog/ai-in-online-advertising-5-key-trends-from-august-2026/">automated bidding</a> as the default for new ad groups. CRM platforms are shifting to pay-per-outcome pricing, with <a href="https://www.hubspot.com/ai-instructions">HubSpot's agent builder</a> charging per resolved conversation and Salesforce following with comparable per-action credits. Adobe put a beta assistant into its video and podcast tools for generating music, speech, and sound effects under its existing commercial-safety terms. On AI search, the numbers keep moving: <a href="https://searchengineland.com/google-zero-click-searches-2026-study-479717">Google's AI answers</a> now show on a quarter to half of queries and push zero-click behavior even higher, while <a href="https://www.searchenginejournal.com/ai-search-citation-sources-trendos-spa/586289/">click-through data</a> shows uncited brands losing ground to cited ones. Referral traffic from these tools is still <a href="https://explodingtopics.com/blog/ai-citation-tracking-tools">small but growing fast</a>, and a new crop of <a href="https://www.conductor.com/academy/best-aeo-geo-tools/">visibility trackers</a> now exist to check whether your own site gets cited.

By now you have two assistants on the bench, and you have probably already caught them disagreeing with each other about your own market. Good. That was the point. This chapter is about the other half of that job: not which tool you ask, but how you ask.

Here is the sentence the whole chapter rests on, and I want to say it plainly before I prove it. The generic slop you get back is almost never the assistant failing. It is a brief with most of the fields left blank, so the tool fills them in with the average of the internet.

Think about what happens when you hand a brief to a freelance writer you have never worked with. If the brief says "write me a blog post about our new integration," you will get something. It will be four hundred words, it will open with a rhetorical question, it will say the word "seamless," and it will be nobody's fault but yours. The writer had no audience, no goal, no length, no sense of how you sound, no idea what you are not allowed to claim, and nothing of yours to look at. They filled the gaps with the safest, most average version of the assignment, because that is the only rational thing to do with a blank field.

An assistant does exactly the same thing, only faster and without the awkward phone call. Every blank field gets filled with the average. And the average of everything ever published on the internet about software integrations is, unavoidably, slop.

So a prompt is not a magic phrase. A prompt is a brief. There are six fields on it, and most people fill in one.

The six are audience, goal, format, voice, constraints, and examples. I am going to walk through them one at a time, and each time I will show you the same job with the field missing and then with the field there, so you can hear what changes.

The job we will use is deliberately small and deliberately real: an email announcing a minor product update. Say you run marketing for a scheduling tool used by small clinics, and the engineers have shipped the ability to send an appointment reminder by text as well as by email. It is not a launch. Nobody is putting it on the homepage. But it needs an email, and that email is sitting in your week right now.

Start with audience. The blank-field version of this is "write an email to our customers." The trouble is that "our customers" is not a person, and the assistant cannot write to a category. It will average across every clinic that has ever bought software, which means it writes to nobody.

Filling the field properly takes three things. First, the one person. Not the segment, the person: the practice manager at a four-chair dental clinic, who does the scheduling herself between patients and is the only one in the building who touches the software. Second, where they are in the buying process. She already pays you. She is not deciding whether to buy, she is deciding whether this new thing is worth ten minutes of her afternoon. That is a completely different job from persuading a stranger. Third, and this is the one people skip, what she already believes. She believes text reminders reduce no-shows because everyone in her industry says so. She also believes that turning on anything new in software means a week of something being broken. So the email is not arguing for text reminders. It is arguing that switching this on will not ruin her Tuesday.

Watch what that does. With the field blank, you get an email that explains why no-shows are costly, which she knows better than you do, and which will make her feel talked down to. With the field filled, you get an email that opens with the fact that it takes one toggle and does not change anything she has already set up. Same tool, same request, entirely different asset. Nothing about that improvement came from clever phrasing. It came from a filled field.

Goal is next, and goal is where most briefs quietly break. A goal is one job. One action, or one decision moved. Not two.

The temptation is enormous. You are writing this email anyway, so surely it can also drive awareness of the wider platform, and pick up a couple of upgrade conversations, and get some replies for the sales team, and remind people the annual conference is coming. Ask for all of that and you will get an email that does none of it. This is not the assistant being lazy. It is arithmetic. Every additional ask takes room from the others, and an email that has four calls to action has effectively none, because the reader satisfies the requirement of "responding" by doing nothing and feeling vaguely informed.

So for our announcement the goal is: she switches text reminders on for her clinic this week. That is it. One action, one week. Everything in the email either moves her toward that toggle or comes out.

Here is the useful side effect. Once the goal is a single action, you can judge the output. You stop reading a draft and asking whether you like it, and start asking whether a busy practice manager who reads only the first two lines would go and flip the switch. Taste is slippery and arguable. That question has an answer.

Format is the third field, and it is the one that costs you the most for the least reason. "Write me a blog post" hands the entire shape of the thing to the tool. Length, structure, whether it has subheadings, whether it opens with a story or a claim, whether the ask sits at the top or the bottom — all of it, decided by the average.

Filling the format field means saying the length, saying the structure, and saying where the thing will actually be read. That last part does real work. Our announcement will be read on a phone, in a gap between patients, in a mail app that shows about forty characters of subject line and one line of preview text. That fact alone dictates that the whole argument has to survive in the first sentence, and that a three-paragraph warm-up is not a stylistic choice, it is a dead email.

So the format line reads something like: one email, subject line under forty characters, one line of preview text, body under a hundred and twenty words, no subheadings, one link, the ask in the first two sentences and repeated at the end. That is not fussy. That is the difference between a draft you edit and a draft you rewrite.

And notice you can be wrong about format and still win. If a hundred and twenty words turns out to be too tight, you now know that, because you set a number and watched what happened to it. When the field is blank, the tool picks a length, you have no opinion about that length, and you learn nothing.

Voice is the fourth, and this is the field people believe they are filling when they are not. Writing "in a friendly, professional tone" fills nothing. Every brand on earth would sign off on friendly and professional. Adjectives about voice are almost worthless, because they describe a target without hitting it, and because every writer means something different by them.

Two things actually work. The first is a few plain sentences describing your own sound in terms of behaviour rather than adjectives. Not "warm and approachable" but: we write short sentences, we use the word "you" and never "users," we never open with a question, we say what the thing does before we say why it matters, and we do not use exclamation marks. Those are checkable. You can hold a draft against them and point at the failures.

The second, and it is stronger, is a pasted sample of your real published copy. Take an email you sent that worked and that sounded like you, paste the whole thing in, and say: this is how we sound, match it. You are giving the tool the actual thing instead of a description of the thing.

Now, an honest note about this field. You are going to type those sentences and paste that sample by hand, and then you are going to do it again tomorrow on a different task, and again the day after. That repetition is real and it is annoying, and it is precisely why the next chapter has you write the voice profile down once as a reusable thing. For now, do it by hand. Feeling the repetition is worth something — it is how you find out which four sentences about your voice actually matter, because those are the ones you keep retyping.

Fifth is constraints, which is the field that keeps you out of trouble. Constraints are the four kinds of "must" and "must not": what has to appear, what must never appear, claims you are not allowed to make, and words your brand does not use. Plus the hard limits on length, if you did not already nail them in format.

For our announcement, that might be: must mention that it works with the existing reminder schedule; must not imply any existing settings change; must not say text reminders reduce no-shows by any specific percentage, because we have no number we can stand behind; do not use the words "seamless," "effortless," "revolutionise," or "excited to announce." One link only, to the setup page.

That claims line deserves a second of attention, because it is the one that protects you. An assistant asked to make a small feature sound worthwhile will reach for a statistic, and if you have not given it one, it will produce a plausible one. A made-up percentage in a customer email is not a style problem. So the constraint is not decoration. It is a fence, and you build it before you need it.

The word list feels petty and is not. Every industry has a handful of words that mark copy as machine-made, and yours has its own. Writing them down as a banned list does more for the finished draft than any amount of instruction about tone, because it is a rule the tool can obey exactly, rather than an aspiration it can only approximate.

Sixth, and strongest of all six, is examples. This is the field with the best return, and it is the one that gets left blank most often.

The move is simple: paste in two or three real pieces of your own good work of the same type, and ask for that pattern. Not "write in our style." Instead: here are three announcement emails we sent that worked, follow this pattern.

Why does that beat describing it? Because a description is lossy and an example is not. When you write "we use short sentences," you have captured one property of your writing out of hundreds. The example carries all of them at once — the sentence length, yes, but also how you open, how you handle the ask, how much you explain before you get to the point, whether you use contractions, how you sign off, the rhythm of the thing. You never had to notice any of that consciously, and you do not have to, because you are not describing the pattern. You are handing it over and letting the tool copy it.

That is the whole trick, and it is worth glossing in one line: the tool is very good at continuing a pattern it can see, and only average at inventing one from adjectives. So show it the pattern.

Two practical notes. Use good examples. Feed it three mediocre emails and it will produce a fourth, faithfully. And use examples of the same job — announcement emails for an announcement email, not your best blog post, because the shape of a blog post is not the shape you want copied.

So: six fields. Now assemble them, because the point of all this is not to reason through six things every morning.

Open a plain text file — a note, a doc, anything you can get to in five seconds. At the top, put the fields that change with the task: audience, goal, format. Underneath, put the fields that barely change at all: your voice sentences, your constraints and banned words, and your examples. Task-specific on top, stable underneath. That ordering is not cosmetic. It means that tomorrow you edit three lines and leave the rest alone, and the bottom half of that file quietly gets better every time you use it.

Then run it. Take the live task you actually have this week and fill the fields out in front of yourself, and do it in one pass rather than in a conversation. One block of text, all six fields, one send. That matters more than it sounds like it does, because the thing you are testing is the brief, and you cannot test a brief you delivered in eleven instalments.

When the draft comes back, judge it against the brief and not against your taste. Go field by field. Is it written to the practice manager or to a category? Does every sentence move her toward the toggle? Is it under a hundred and twenty words with the ask in the first two sentences? Does it sound like the sample you pasted? Did it obey the fence — no invented statistic, none of the banned words? Does it follow the pattern of the three examples?

You will usually find that it fails on one or two of those and passes the rest. And now comes the single most valuable habit in this chapter, which is what you do about it.

Do not fix the output. Fix the field.

If the draft is too long, your format field was thin. If it sounds like a press release, your voice field was thin, or you gave it adjectives instead of a sample. If it invented a number, your constraints field had no fence in it. If it explained why no-shows are costly, your audience field never said what she already believes. Each failure points at exactly one field, and repairing that field is a change that survives. Rewriting the output is a change that dies the moment you close the tab.

Which brings us to the pitfall, and it is the one you will hit this week if nobody warns you.

You get a draft that is almost right. Ninety percent. So you reply: shorter. It gets shorter and loses the ask. You reply: put the ask back. It comes back with the ask and an exclamation mark. You reply: no exclamation marks, and warmer. It gets warmer and starts calling her a user. Ten exchanges later you have something worse than the first draft, you are annoyed, and you have spent forty minutes on a hundred-word email.

That is the fix-the-output trap, and the reason it produces mush is structural. Every correction you type lives only inside that one conversation, competing with everything else you have said in it. You already know from picking your assistants that long chats drift as the early instructions lose ground to the recent ones. Arguing with a draft is the fastest way to build exactly that kind of overloaded, drifting conversation. And when you finally get something usable, none of it is saved. Tomorrow's email starts from the same blank fields.

The tell is precise and easy to catch: you are giving the same note twice. The second time you type "shorter," or the second time you have to say you do not use exclamation marks, stop. Not because the assistant is being stubborn, but because a note you have to give twice is a missing field, and you have just been handed its exact contents for free.

So take the note out of the conversation and put it in the file. "No exclamation marks" goes into constraints. "Never opens by explaining the customer's own problem to them" goes into voice, or better, gets solved by pasting a better example. Then start a fresh run with the improved brief. You will often find the second run needs no follow-up at all, and it took less time than the argument would have.

That is the whole loop, and it is boringly mechanical, which is the point. Brief, run, judge against the brief, find the thin field, fix the field, run again.

Your next action is small. Take the one real task you have waiting — the announcement, the ad, the caption, whatever is actually on your list — and fill in the six fields in a file you can paste from. Run it once, whole. Then take whatever the first correction was, the very first note you found yourself wanting to give, and write it into the constraints line before you close the file. That is it. One live task, one correction banked. Your template goes into next week one field stronger than it came in, and that bottom half — the voice, the constraints, the examples — is the raw material for the profile you will build next.

Now, the news, because some of what shipped this month changes what a brief is even for.

The clearest one is that the ad platforms are taking more of the decisions off your desk. In the middle of August, the advertising side of ChatGPT switched on an automated bidding option — the sort that sets and adjusts your bids for you inside a budget, aimed at whatever objective you picked — and made it the default for new qualifying ad groups. The documentation is careful to say it guarantees no particular cost per acquisition, cost per click, or return on ad spend. Why you care: the default has changed, which means it applies whether you chose it or not. Smallest next action is to open any ad group you have created recently and confirm which bidding method it is actually using.

The bigger structural story is that the customer relationship management suites — the systems of record for your contacts and deals — are now selling agents by the outcome rather than the seat. HubSpot's agent product went into public beta over the summer and added a builder in August that lets you assemble your own agents by describing them in plain language and hooking them to triggers in the CRM. Alongside that, its pricing moved to charging per result: reportedly around fifty cents per resolved customer conversation, a dollar per recommended outreach lead, and about ten cents a query for data answers. Salesforce went the same direction inside its marketing cloud across July and August, adding pay-per-resolution billing next to its credit packs, which work out at roughly ten cents an action, with per-user add-on licensing starting around a hundred and twenty-five dollars a month and the marketing edition that includes campaign generation listed at fifteen hundred a month per organisation, billed annually. Why you care: this is a cost-math change, not a feature change. When you pay per resolved thing, volume is your budget line, and you can forecast it. Smallest next action is to work out what one resolved conversation is currently worth to you in labour, and compare.

On the creative side, Adobe put an assistant into beta in late August that generates music, speech and sound effects for video and podcast work, inside its commercial-safety guarantees. That last part is the reason it earns a mention rather than the audio itself.

And the standing check on AI search, which is a moving picture rather than a settled fact. Google's AI answers now appear on somewhere between a quarter and a half of queries depending on market and who is measuring. Zero-click rates across web search sit somewhere in the high fifties to high sixties percent overall, rising to roughly eighty-three percent when an AI answer is shown, and reportedly as high as ninety-three percent inside Google's conversational mode. The top organic result loses something like a third to a half of its click-through when an AI answer appears — but brands that get cited in the answer pick up meaningfully more clicks than uncited competitors, on both the organic and paid side.

The referral traffic itself is still tiny, on the order of a tenth of a percent to about one percent of all web traffic, though growing at several hundred percent year over year, and ChatGPT's share of it has fallen from the high eighties to around sixty-three percent as Claude, Gemini and Perplexity take ground. Those sessions do convert better than ordinary organic. The number that should change your content plan is this one: across engines, earned and third-party media account for roughly eighty-four percent of citations, while your own website typically supplies only five to ten percent of cited links, and community sites like Reddit and YouTube make up around half. Meanwhile more than half of the pages cited in Google's AI answers are not in the traditional top ten at all.

Tracking this has become its own small tool category, with monthly prices running from about twenty dollars at the low end to a few hundred, and the big SEO platforms have added visibility modules of their own, at least one with a free entry tier. Smallest next action: pick one tracker with a free or cheap tier, put in the five questions a real prospect asks right before buying, and see whether you are named. Whatever you find, the workflow is the thing that lasts — the trackers will churn, and the question of whether you are the cited source will not.