The short version: a truly awesome ChatGPT prompt gives the model a role, a real piece of context, a specific job, and a clear format for the answer. That’s it. No secret phrases, no “act as the world’s greatest expert” nonsense, no magic words that unlock some hidden mode. I’ve written thousands of these things since 2023, tested them on my own business and on client projects, and the pattern is boringly consistent.
The bit nobody wants to hear about prompt engineering
There are courses out there charging £300, £500, even £1,200 for “prompt engineering mastery.” I’ve been sent dozens of them to review. Most of them are teaching you formulas that were true for GPT-3.5 in early 2023 and are already half wrong for the model you’re using right now. ChatGPT gets updated constantly, the underlying model changes, and a “trick” that worked brilliantly in March might do nothing by September. I tested one popular formula from a paid course in October 2025 against the same task without the formula and got near identical output both times.
That’s the uncomfortable bit. A huge chunk of what’s sold as prompt engineering isn’t engineering at all, it’s pattern matching on a moving target. The people making real money from prompts aren’t the ones with the fanciest templates, they’re the ones who know exactly what result they need before they open the chat window. The prompt is just the delivery mechanism for that clarity. If you’re vague in your own head, no template fixes that.
What an awesome prompt contains
Strip away the hype and every prompt that has ever worked well for me has four things in it. Miss one and the output gets generic, waffly, or just wrong.
- A role or lens. Not “act as an expert,” that’s meaningless filler. Something specific, like “you’re a B2B copywriter who writes for busy operations directors who hate jargon.”
- Real context. Actual detail from your world, not hypotheticals. Names, numbers, previous attempts, what’s already been tried and failed.
- A single, specific job. Not “help me with marketing,” but “write three subject lines for an email announcing a price rise to existing customers.”
- A format for the output. Length, tone, structure. Bullet points or prose. 100 words or 500. A table or a list.
That’s four sentences, sometimes five. It doesn’t need to be a page long. In fact the prompts I see clients over-write, the ones stuffed with ten paragraphs of backstory, tend to perform worse than a tight four-line version, because the model has to hunt through the noise to find the actual instruction.
A real example from my own business
Back when I was rebuilding my consultancy after a rough few years, I needed to write a sales page for a workshop I was launching. My first attempt at a prompt was something like: “Write a sales page for my AI workshop, make it persuasive and engaging.” What I got back was a slab of generic copy that could have been for anyone’s workshop about anything. Fine grammar, zero soul, nothing I could use.
Second attempt, I gave ChatGPT this instead: “You’re writing for me, Lilach Bullock, a 53 year old AI and marketing consultant who was on the Forbes list, ex influencer, now rebuilding my business publicly. The audience is small business owners who are curious about AI but scared of looking stupid using it. The workshop is three hours, live, £197, covers using ChatGPT for content and lead generation, no fluff, no coding. Write a 400 word sales page. Open with a specific, uncomfortable problem this audience has right now, not a generic hook. End with one clear call to action.”
The difference was night and day. The second version had a real opening line about people copying and pasting ChatGPT answers straight into client emails and getting caught out. That came from the specificity I gave it about the audience’s fear, not from clever wording. I didn’t change the “style” of the prompt, I changed how much of my actual knowledge was in it. That workshop sold out its first 20 spots inside four days, and I still use variations of that same sales page structure now.
The step by step, if you want a repeatable process
Here’s the exact sequence I run through, whether I’m writing a prompt for myself or teaching it to a client on a one to one AI coaching call.
- Write down the job in one sentence before you open ChatGPT. “I need a LinkedIn post about why my client’s revenue dropped and what I’m doing differently.” If you can’t do this, stop, you’re not ready to prompt, you’re ready to think.
- Add the role. “You’re a LinkedIn ghostwriter who writes short, punchy, first person posts for consultants.”
- Dump in the raw material. Paste actual notes, numbers, quotes, whatever real detail exists. Don’t summarise it first, let the model do that work.
- Set the constraint. Word count, tone, what to avoid. “Under 150 words, no hashtags, no emojis, one clear point.”
- Ask for three versions, not one. This single step improves my output more than almost anything else. One version gives you the model’s default. Three gives you a choice, and often the third one, the one that took the biggest risk, is the best.
- Edit it like it’s a junior copywriter’s draft, not gospel. Every single output needs your voice run through it. If you’re not changing at least 20 percent of the words, you’re not editing enough.
That process takes me under five minutes for most short content and produces something usable, not perfect but usable, on the first or second pass.
Where people go wrong most often
I read a lot of prompts, from clients, from people in my community, from strangers on LinkedIn asking me to look at what they typed. The mistakes repeat.
- Asking for too much in one go. “Write me a full content strategy, three blog posts, five social captions and an email sequence” in a single prompt. You’ll get shallow everything instead of good something. Break it into separate, focused requests.
- No audience specified. “Write a blog post about time management” gives you an answer for everyone, which means it’s useful to nobody. “Write for solo parents running a small online shop” gives you something with an edge.
- Treating the first answer as final. The first response is a draft, not a decision. I almost never keep the first output for anything client facing.
- Forgetting to give it your existing material. If you’ve already written ten emails to this client, paste one in and say “match this tone.” Don’t make the model guess your voice from nothing.
- Over-formatting the request itself. Bullet pointing your prompt into seventeen numbered rules can backfire, the model starts answering the structure instead of the actual question. Plain sentences with the four elements above usually beat a checklist-style prompt.
Why the “perfect prompt” you save and reuse will stop working
This is the part that annoys people who’ve bought a £30 prompt pack from a marketplace. Those saved prompts age badly. A prompt that gave you gold in June might give you something flatter in December, not because you did anything wrong, but because the model itself shifted underneath you, the training data updated, the guardrails changed, the default tone moved. According to Wikipedia’s overview of prompt engineering, the practice is explicitly tied to how a specific model interprets language at a specific point, which is a polite academic way of saying: what works today is not guaranteed to work in six months.
That’s not a reason to give up on having a personal prompt library, I keep one myself, about 40 prompts I reuse and tweak constantly. But it is a reason to treat every prompt as a living draft, not a spell you cast once and trust forever. Check your output quality every few weeks. If something that used to sing starts sounding flat, don’t blame yourself, rewrite it fresh with more current context.
The one thing that matters more than wording
If I had to strip this whole topic down to a single sentence, it’s this: the quality of your prompt is a direct reflection of the quality of your thinking, not your vocabulary. I’ve watched people with beautifully worded, grammatically perfect prompts get mediocre output because they hadn’t decided what they wanted the piece to achieve. And I’ve watched people with clumsy, typo-ridden prompts get brilliant output because they knew precisely what problem they were solving and who for.
So before you touch the keyboard, answer three questions in your head: who is this for, what should they do or feel after reading it, and what’s the one fact or detail only I would know to include. Get those three answered and the actual prompt writes itself in under a minute.
If you’re running a small business and this whole area feels like more time than you have, that’s exactly the kind of practical, hands-on work I cover with clients through AI implementation coaching, building prompts and workflows around your actual business instead of generic templates.
Frequently asked questions
Does the order of information in a ChatGPT prompt matter?
Yes, to a degree. Putting the role and the specific task early tends to anchor the response better than burying it at the end of a long paragraph, but it matters far less than simply including all four elements: role, context, task and format. Order is a small tweak, presence of information is the real driver of quality.
Are longer prompts better than short ones?
No. Length should match the complexity of the task, not the other way around. A one-line prompt with real specificity beats a ten-paragraph prompt full of vague throat-clearing. I’ve had strong results from four-sentence prompts and weak results from prompts three times that length that never said what the person wanted.
Should I use special phrases like “think step by step” or “you are the world’s best expert”?
Some of these phrases had a measurable effect on older models and have far less impact now that models reason more consistently by default. Rather than chasing magic phrases, focus on giving genuine detail about the audience, the goal and the constraints. That’s what still reliably improves output.
How many times should I edit a ChatGPT prompt before giving up on it?
If three attempts with clear, specific instructions still produce weak results, the problem usually isn’t the wording, it’s that you haven’t decided the actual outcome you want. Step away from the chat, write one sentence describing the goal on paper, then go back and prompt again. That single step fixes most stuck prompts faster than another round of rewording.
The same rule explains: why most funny ChatGPT prompts fall flat.
Related reading: ChatGPT Prompts for Parents to Keep Kids Busy (Without the Guilt).
