If AI writing reads like AI writing, the model is usually not the problem. The prompt asked for the average of everything ever written on the topic, and that is exactly what came back.
Marketers have been briefing freelancers for decades and know instinctively that "write me a blog post about our product" produces something unusable. The same instinct applies here and is routinely forgotten, because a chat box does not look like a brief. It is one.
What follows are the five techniques that move output from generic to usable, each with the before-and-after. None of them require technical knowledge. All of them are about supplying decisions the model otherwise has to guess.
1. Specify the reader, not the topic
A topic tells the model what to write about. A reader tells it what to leave out - which is the harder and more valuable instruction.
Before: "Write a blog post about marketing automation."
After: "Write for a VP of Marketing at a 40-person B2B software company who has been burned by an agency retainer that produced activity reports instead of pipeline. They are sceptical of AI tools because they have tried three that produced content they had to rewrite. They are not looking to be educated about what automation is."
The second version cannot produce a "what is marketing automation" introduction, because the reader has been defined as someone who would be insulted by one. Most generic output is generic because nothing in the prompt ruled anything out.
The practical test: does your reader description contain something that makes an obvious section inappropriate? If not, you have described a demographic, not a reader.
2. Give it raw material it could not have
A model's default output is a synthesis of what is publicly common. If you supply nothing proprietary, you get the consensus - competently expressed, and indistinguishable from what your competitor gets.
Before: "Write a case study about a customer who improved their conversion rate."
After: "Write a case study from these notes. Customer: regional accounting firm, 12 staff. Before: 4 inbound consultations per quarter, 61% of calls unanswered during business hours, no after-hours cover. Changed: inbound capture and qualification automated; booking moved onto the call itself. After, over two quarters: 19 consultations per quarter, unanswered calls under 5%. Quote from the managing partner: 'We weren't losing to competitors on price. We were losing to whoever picked up.' Do not round the numbers and do not add any metric that is not in these notes."
The second prompt produces something only you could publish. It is also far less likely to invent a statistic, because the closing instruction removes the excuse for one - though "less likely" is the honest phrasing, not "cannot". Check the numbers anyway; that is what the review step is for.
This is the single highest-leverage habit available to a marketer, and it costs nothing except keeping notes.
3. Constrain the shape, not just the length
"Write 800 words" constrains volume, which is the least interesting property of a piece of writing. Structure is what makes output usable or throwaway.
Before: "Write a 600-word LinkedIn post about our new feature."
After: "Open with the specific moment the problem shows up in someone's week - no statistics in the first line. Then three short paragraphs: what people currently do, why that fails at scale, what changes. Close with a question that only someone with the problem would answer. No emoji, no bullet lists, no 'Here's the thing:'. Under 200 words."
Named prohibitions do real work. Models have strong stylistic defaults, and the fastest route to a distinctive voice is often to enumerate the defaults you refuse.
4. Show the voice instead of describing it
"Professional but friendly" means nothing. Every brand believes it is professional but friendly, and the phrase constrains nothing.
Before: "Use our brand voice: professional, approachable, slightly irreverent."
After: "Match the voice of this paragraph from our site - [paste 100 words of your actual best writing]. Specifically: sentences run short, we use 'you' and never 'users', we state the objection before answering it, and we never open with a statistic."
Two or three hundred words of your own strongest copy will out-perform any adjective list. If you cannot find three hundred words of your own writing you would be happy to imitate, that is a useful finding in itself - and the real task is writing them, not prompting around their absence.
5. Ask for the thinking before the output
The costliest failure mode is a beautifully written piece built on the wrong strategic premise. You then edit prose when you should have rejected the angle.
Before: "Write a landing page for our Pipeline Team product."
After: "Before writing anything, give me: the single most likely objection this buyer raises in the first ten seconds, the one claim the page must prove to overcome it, and the proof we would need. Wait for my response before drafting."
This is the technique marketers adopt last and value most. It surfaces the disagreement at the cheapest possible moment - before the draft exists and before anyone is attached to it.
Four habits that quietly ruin output
- Politeness padding. "Could you please help me write..." adds nothing. Instructions are not rude.
- Stacking unrelated asks. One request per prompt. A prompt that asks for a blog post, five social variants, and an email sequence produces three mediocre things instead of one good one.
- Superlative instructions. "Make it engaging and compelling" is unfalsifiable, so it changes nothing. "Cut every sentence that a competitor could also have written" is testable, so it does.
- Editing the output instead of the prompt. If you fix the same problem by hand twice, the fix belongs in the prompt. Otherwise you are doing manual labour indefinitely and calling it automation.
The shift that actually matters
Every technique above is the same move: supply a decision the model would otherwise make for you. Who this is for. What it must not say. What shape it takes. What it sounds like. What it must prove.
Which means prompt engineering is not a new discipline for marketers. It is the brief - the thing good marketers were already writing and often skipping. The novelty is that the consequence of skipping it now arrives in four seconds instead of four days, and looks polished enough that you might publish it.
This is also why "the AI wrote something generic" is rarely a useful complaint. It is almost always a report about the brief.
Where to start
Take the last piece of AI output you rejected. Work out which of the five decisions above the prompt left to the model. Add it. Run it again. Most of the time, one addition is the difference.
Our own working prompts - the ones the crew runs for audits, campaigns, content, pipeline, and reporting - are published in the prompt library, built to the structure described here. They are worth reading for the shape as much as the content: every one of them names the reader, supplies the raw material, and states what the output must not do.
Find your next marketing priority
Start for free with a public-site audit, then choose Growth or Scale to run marketing outcomes with the crew.
Run free audit