You ask for a blog post about email marketing. Twenty seconds later, 600 words arrive: cheerful, balanced, and completely unusable. No argument. No numbers. No opinion. Just the statistical average of everything ever published on the topic.
The model isn’t the problem. The prompt was written like a note to a colleague who already knows your business, your reader, and what happens to the output next. ChatGPT knows none of that. A decent ChatGPT prompt closes the gap in four or five lines.
What a prompt is actually doing under the hood
The model predicts plausible text. Give it a vague scenario and the most plausible continuation is the blandest one, because blandness is the safest prediction. Constraints shrink that space. Name an audience, a length, a stance, a format, and the plausible answers get pushed somewhere specific.
This explains a pattern you’ve probably noticed. “Be more creative” changes almost nothing. “Write this as a complaint letter from a customer who has been waiting 11 days for a refund” changes everything, because it hands the model a voice and a grievance to work from.
The four parts that carry most of the weight
Nearly every prompt I reuse follows the same skeleton. Skip one part and you’ll feel it in the response.
- Who is speaking, and to whom. “You’re a commercial property lawyer explaining a break clause to a first-time tenant” beats “you’re a lawyer” every time. The reader matters as much as the role.
- The task, stated as a deliverable. Not “help me with pricing” but “write three pricing page variants for a £29/month tool, each under 90 words.”
- Constraints with real numbers. Word counts, reading level, number of options, words to avoid, facts to include. Numbers are the cheapest way to steer output.
- The shape of the answer. A table, five bullets, three paragraphs, a 40-word subject line. If you don’t specify, you’ll get a wall of prose with a tidy intro.
If you only fix one thing, fix the third. Vague prompts produce vague answers, and specificity is the whole game, as this walkthrough on how to stop getting generic answers from ChatGPT makes clear with side-by-side examples.
Prompts worth copying today
These are short, and each one does a job I’d otherwise do badly at 9am.
For cutting, not writing
“Here’s 400 words of my draft. Cut it to 180 without losing the argument. Delete anything that restates the previous sentence. Return only the new version, then a two-line note on what you removed.”
For pressure-testing an idea
“I’m going to describe a plan. Give me the three strongest objections a sceptical finance director would raise, ranked by how likely they are to kill the project. Be blunt. No encouragement.”
For turning mess into structure
“Below are nine scattered notes from a client call. Group them into themes, name each theme in four words or fewer, and flag anything that sounds like an unresolved decision. Use a table.”
For the awkward email
“Write a 90-word email to a supplier who has missed two deadlines. Firm, not aggressive. No exclamation marks, no ‘I hope you’re well.’ Offer one concrete next step and a date.”
For planning something real
Prompts get surprisingly good when you force the model to work with actual constraints. A full eight-step plan for a street party, including budget lines and a rainy-day contingency, shows how far a structured chain of prompts can go, which is the approach set out in planning anything with a ChatGPT AI chatbot.
Editing prompts outperform writing prompts
Most people use ChatGPT to generate. The better return comes from using it to revise text you already wrote, because then the judgment is yours and the model just does the labour.
Two habits help. Paste the draft rather than describing it. And give the model a target to optimise against: “reads like a trade magazine,” “fits in a LinkedIn post without a ‘see more’ cutoff,” “could be read aloud in 45 seconds.” Without a target, editing prompts drift into rewriting your voice into a neutral one.
Chain prompts instead of writing one enormous one
The instinct is to cram everything into a single mega-prompt with eleven requirements. It usually underperforms, because the model juggles and drops things. Three or four sequential prompts, each with one job, are more reliable and easier to debug.
A workable sequence looks like this: extract the raw material from rough notes, restructure it, draft it, then edit it against a checklist. Each step sees the previous output and nothing else, which keeps it focused. There’s a full worked version of this in a five-step workflow that turns rough notes into finished work, and the difference between chaining and one-shot prompting is genuinely large.
One practical tip: keep each prompt in the chain under about 120 words. Long prompts invite the model to satisfy the easy requirements and quietly ignore the fiddly ones.
Where prompts still fall over
Some jobs resist prompting no matter how carefully you write. Be sceptical of anything involving numbers you haven’t supplied, recent events, or sources you can’t check. Ask for a calculation and you may get confident arithmetic that’s wrong by a factor of ten. Ask for a citation and you may get a real-sounding title attached to an author who never wrote it.
The fix isn’t a cleverer prompt. It’s changing what you ask for: request the method rather than the answer, or paste in the figures and ask the model to work only from those. Prompts steer language brilliantly and facts unreliably. Treat them accordingly. If you’re weighing up which plan or model to run them on, this breakdown of what ChatGPT Pro actually buys you for $200 a month is a useful reality check before upgrading.
Keep the prompts that worked
After a few weeks of this, you’ll notice maybe eight or nine prompts doing 80% of the useful work. Write them down somewhere you’ll actually look. A plain text file is fine. Name them by job, not by topic: “cut draft,” “client call notes,” “supplier email.”
Then improve them one constraint at a time. If the output is too long, add a word count. If it’s too formal, name a publication whose tone you want. If it keeps adding a summary paragraph you never asked for, say so explicitly. Small edits compound, and after a month you’ll have a set of prompts tuned to your work that nobody else could have written. The six-step method in getting real work out of ChatGPT 4, with actual prompts is a good place to borrow structure from while you build that list.
The habit that beats every template: before you hit enter, read your prompt back and ask whether a competent freelancer who has never met you could produce the right thing from it. If they’d have to guess at the audience, the length, or the format, so would the model. Add the missing line. That single check is where most of the improvement lives.