The short version: Tell AI what format you want, who it's for, and what success looks like. Generic prompts give generic output; the difference between "write a social post" and "write a LinkedIn post for UK finance directors about cash flow forecasting, 280 characters, no jargon" is the difference between bin and broadcast.
Why your prompts are failing
Last month I asked ChatGPT to write email subject lines for a B2B SaaS product launch. It gave me five options that sounded like they'd been written in 2008. All caps, exclamation marks, nothing that would make a single person click.
Then I rewrote the prompt with three changes: I specified the audience (already-paying customers aged 40-55), the pain point (upgrade fatigue), and the tone (conversational, no hype). The output shifted immediately. Usable. Not genius, but sendable.
The problem is that AI models don't read minds. They read instructions. Most marketing people ask them questions. That's not how this works.
You're not asking. You're building a specification document. The more detail, the better the output. This is not intuitive if you've spent years working with human writers who can infer context.
The four things every prompt needs
1. The actual format, specific and measurable
Not: "Write copy for our homepage."
Yes: "Write the above-the-fold headline and 40-word subheadline for a homepage targeting UK small business accountants who use Excel and Xero. Tone: reassuring, not fancy. No mention of 'solutions' or 'transformation'. Avoid all sales language."
When you say format, I mean: how many words, how many sentences, how many bullet points, what line breaks, what structure. ChatGPT can count. Use numbers.
I was working with a fashion e-commerce founder last week who needed product descriptions for 200 items. Her first batch from the AI were all 90 words each, flowery, and identical in rhythm. I rewrote the prompt to specify: "50-60 words. Three sentences max. First sentence is what it is. Second sentence is who wears it. Third sentence is why it matters. No adjectives stacked together."
The next batch was salvageable. Not perfect, but 70% had no rewrites needed.
2. The audience, described like you're briefing a copywriter
Not: "Write for our customers."
Yes: "Write for Karen, 52, runs a 12-person marketing consultancy from home in Manchester, uses Slack and Monday.com, spends between 3pm and 10pm on client work because mornings are admin, frustrated with how long it takes to brief junior staff, makes 80k a year and reinvests 40% of it back into training."
The more specific you get about who reads this, the more the AI can predict what language will land. Include age, job title, location if it matters, tools they use, their frustration or goal, and their budget bracket if relevant.
I spent two years as an influencer before building my marketing business. I learned early that vague briefs produce vague work. It's the same with AI. The AI doesn't know who this is for unless you tell it with detail that would look excessive in a brief to a human copywriter.
3. The context or constraint that makes this different
Not: "Write a LinkedIn post."
Yes: "Write a LinkedIn post about why businesses should hire freelance marketers instead of hiring in-house. Our competitor is HubSpot's hiring services. We're cheaper but less shiny. Our audience has already rejected us once. Keep it under 150 words. No emoji. Don't pretend we're better than we are. Tone: honest, slightly self-deprecating."
The constraint is often what makes the work good. Without it, AI will just say what everyone else says. With it, you get something specific that fits your situation.
Last week I needed copy for a workshop I'm running in March. I told the prompt: "You're writing a sales page for a workshop that teaches people how to audit their AI spending. Most of the audience have over-invested in tools they don't use. The workshop costs 297 pounds. The outcome is they'll save money, not make money. They're tired of hype. Write a 100-word landing page section that doesn't try to upsell them on AI. Make them believe they'll get their money back."
That last constraint changed everything. The AI stopped trying to make it sound exciting and started making it sound honest.
4. What counts as success or what to avoid
Not: "Make it good."
Yes: "Success = we can send this without rewrites. Avoid: jargon (no 'solutions', 'synergy', 'innovative'), ALL CAPS, rhetorical questions, more than two exclamation marks, mentioning competitors by name, anything that sounds corporate."
Tell the AI what you're trying to avoid. That's often more useful than telling it what you want. Instead of "make it casual", say "avoid corporate language, no use of 'we' three times in a row, no sentences starting with a noun that describes a feeling (excitement, opportunity, etc.)."
I also specify: "If you can't do this in one draft, tell me what's blocking you." That stops the AI from generating garbage and instead making me work with it on what's hard about the brief.
How to structure your prompt for maximum usability
The order matters because the AI processes early information more carefully.
Start with what you want: "Write a 60-word Facebook ad for UK women over 40 interested in strength training."
Then add who: "The audience has tried fitness before, usually hated it, thinks they're too old. They're not Instagram people. They're WhatsApp people. They make 50-100k a year. They want results, not community."
Then add the constraint: "Our competitor is F45. We're a home-based program, not a studio. We cost half as much. We're slower but more sustainable."
Then add success criteria: "Success = they believe they can start this week. Avoid: before-and-after language, 'transform your body', anything motivational, talking about age as a barrier."
Then, if you have examples, add them: "Here's one we sent last month that worked: [paste an ad you ran]. Do something in that direction but different."
Then the instruction: "Generate three versions. For each version, tell me which part you're changing from the example and why."
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
This structure takes 45 seconds to write and saves you hours of rewrites.
Real examples that worked
I ran a series for a B2B event in October. Here's the prompt I used for LinkedIn carousel ads:
"Create a 5-slide LinkedIn carousel ad selling tickets to a half-day AI workshop for freelance marketers. The workshop costs 247 pounds. Audience: UK freelancers, 35-55, been freelance for 5-15 years, made between 40k-120k last year, curious about AI but skeptical about hype, don't have time for 'learning journeys'. Success = they believe they'll learn something specific that makes their client work faster. Each slide should be 15-25 words max. Slide 1 is the hook (what they'll learn, not why AI is good). Slides 2-4 are three specific things covered in the workshop. Slide 5 is the CTA. Avoid: 'transform', 'unlock potential', 'the future', 'cutting edge', 'empower'. Show me all five slides in a list."
I ran that output with one round of edits (changing "streamline" to "speed up"). We got 340 clicks from that ad. My previous version, written without this level of specificity, got 80.
Another one: I needed subject lines for an email sequence to people who'd bounced from a masterclass. Here's what I used:
"Write 7 subject lines for a re-engagement email sequence. Recipients are UK business owners aged 40-60 who registered for a free masterclass on AI ROI but didn't show up. They're not flaky, they're just busy. The first three emails are: 1) here's the recording, 2) here's the free resource we gave attendees, 3) here's a discount if you book a 1-1. Make each subject line seem like it's from a real person, not a sales machine. No emoji. 20-50 characters each. Success = they feel like we're not going to guilt them. Avoid: 'did you miss us', 'we've been thinking about you', 'last chance', anything guilt-based."
Four of those seven went into live emails. They pulled a 28% open rate on a 40% unopen list.
The difference was time spent in the prompt, not genius in the AI.
How much detail is too much
The honest answer: I've never hit "too much" with ChatGPT 4 or Claude 3.5. The longer and more specific the prompt, the better the output.
If you've written more than a page, you're probably overthinking it. But a detailed paragraph is the minimum.
I usually spend 15 minutes on a prompt that I'll only use once, and 20 minutes on a prompt I'll use repeatedly (because I might as well save the good ones for next quarter).
That's not a lot of time for output you won't need to rewrite.
What to do when the output is still bad
If you get rubbish back, the prompt is the problem, not the AI. Here's my diagnostic:
- Read what you got. Write down the three worst parts.
- For each bad part, ask: did I tell the AI not to do this? If no, add it to the prompt under "Avoid".
- Run the same prompt again. Usually it fixes it on attempt two.
- If it doesn't, the constraint itself is the problem. You've asked for something that conflicts with something else in the brief. Simplify. Cut the brief in half and ask for two separate pieces instead.
I once asked an AI to write sales copy that was "honest but compelling and doesn't mention price." The output was contradictory nonsense because I'd asked for something that can't exist. When I split it into two jobs (one for the landing page without price, one for the email that goes after they clicked), both came back usable.
The single thing that changed my AI output quality
Stopping using the word "write" and starting to use "create", "draft", "produce", or "generate".
I know that sounds silly. It's not. When I say "write a subject line", the AI thinks I want prose. When I say "generate three subject line options with notes on why each one might work", the AI treats it like a working brief and the output is immediately more useful.
Try it. Swap out "write" for "draft" or "create" and watch the quality shift.
The Prompt Structure I Use for Client Content (With Real Numbers)
I ran a small test last year across 40 client prompts to see what changed output quality, and the biggest factor was not the prompt length or the tool, it was whether I included a rejected example. When I gave the model one paragraph I had already written and thrown away because it was too generic, alongside a note saying "do not write like this, here is why it fails," acceptance rates on first draft jumped from roughly 30% to about 70% based on my own tracking sheet. Most advice tells you to give a good example. Almost nobody tells you the bad example does more work.
Here is the exact structure I now use for a LinkedIn post prompt:
- Context: who the audience is, in one sentence (example: "small business owners with under 10 employees who are skeptical of AI tools")
- Bad example: a real paragraph that sounds like every other post on the topic, with a one line note on why it fails
- Good example: 3 to 4 sentences in the actual client voice, pulled from a past newsletter or transcript
- Constraint: a hard number, like "under 120 words" or "exactly one question at the end"
- Output instruction: what to leave out, not just what to include, such as "no rhetorical questions in the first line"
The constraint line matters more than people think. I used to ask for "a short post" and got anything from 60 to 400 words back. Once I switched to giving an exact word count, variance dropped to within about 15 words either side, which is close enough to publish without a rewrite. I also stopped asking models to "sound conversational" because that phrase produces the same fake-casual tone every time, full of phrases like "let's dive in." Instead I ask it to write the way someone would explain the idea to a colleague over coffee who already knows the industry, which produces noticeably fewer filler phrases.
One honest caveat: this setup works well for repeatable formats like LinkedIn posts, email subject lines, and ad variations. It works far less well for long-form blog drafts, where the model still tends to pad sections regardless of how tight the prompt is. For anything over 800 words, I use AI for the outline and first draft of individual sections separately, never the whole piece in one prompt.
Frequently asked questions
Should I be specific about tone, or is that implied?
Never implied. Tell the AI exactly. "Professional but not stiff", "casual without being unprofessional", "warm but not false", "direct without being rude". If you can't describe it, the AI can't hit it. Tone is usually the first thing that's wrong in unusable output, which means you under-specified it.
What if the output is technically correct but sounds like AI?
You didn't specify voice. Add a line: "Write this in the voice of [specific person], not a marketing department." Or: "Remove every phrase that sounds like a press release. Common ones: 'excited to announce', 'thrilled to share', 'we are committed to', 'cutting-edge', 'synergy'." Be explicit about AI-speak you hate.
How do I know if my prompt is good before I run it?
Read it out loud. If you'd never brief a human copywriter this way, the AI won't understand it either. If you found yourself using vague words like "good" or "professional" or "engaging", add specific examples instead: "like the tone of Sathya Alagappan's writing, not like HubSpot's blog".
Can I use the same prompt twice and get the same output?
No. Each run produces different output even with identical prompts. If you get something perfect, save it. If you need consistency (same tone across 50 product descriptions), add the successful output to your next prompt as an example and tell the AI "match this style".
Related reading: How to Use AI for SEO Without Getting Penalised by Google and How to Stop AI Writing from Sounding Like AI.
