A good prompt tells the AI four things: who it is, who you are, what you need and what shape the answer should take. That’s it. Miss those four and you get the generic sludge that convinces half of small business that AI is overrated. Nail them and the same free tool starts producing work you’d actually send. The tool was never the variable. The ask is.
I’ve used the same line in workshops for years: prompting is like giving directions to a mate who’s never been to your place. Say get here quick and they’ll end up lost. Say turn left at the servo, then second right past the big gum tree, and they land on your doorstep with a smile. Same brain, better directions.
The four-part ask
Here’s the template I teach in every session. Save it somewhere you can reach:
You are a [role]. I am a [who you are]. I need [the task] in a [tone]. Return it as [format].
- You are a [role]. Gives the AI a job. A world-class bookkeeper answers differently to a marketing coach. No role means it averages across everything, and average is exactly how the output reads.
- I am a [who you are]. Your context. A gym owner specialising in post-surgery rehab gets rehab-shaped answers. Leave this out and the AI is guessing your entire situation.
- I need [the task] in a [tone]. One task, stated plainly, with the voice you want. Warm and casual is a different email to formal and brief, and the AI can do either. It just can’t read your mind.
- Return it as [format]. Email, table, dot points, 100 words, three options. Constraints are a gift; without them, AI tools ramble like a bloke at a barbecue who’s just discovered crypto.
Three before-and-afters
The difference isn’t subtle. Same tool, same day, different ask.
The social post.Before: write a social post about my cafe. You’ll get something with a coffee emoji that could be any cafe on earth. After: You are a social media writer for local hospitality businesses. I am the owner of a beachside cafe in northern NSW known for sourdough and dog-friendly tables. I need three Facebook post drafts about our new winter menu in a warm, cheeky local voice. Return each under 60 words with no hashtags.
The welcome email. Before: write a welcome email for new clients. After: You are a business development writer. I am a gym owner specialising in post-surgery rehab. I need a welcome email for new clients in a warm, casual tone that covers our opening hours, socials and best contact details. Return it as plain text under 150 words. That second version has been earning its keep in my workshops since the first time I wrote it.
The decision summary. Before: summarise this document. After: You are an analyst who writes for busy owners. I am deciding whether to renew this supplier contract. I need the three things in this document that should change my decision, in plain language. Return them as dot points with a one-line recommendation. Now the summary works for you instead of just being shorter.
Why does the same prompt give different answers?
Because AI is a probability engine, not a rules engine. Automation does the same thing the same way every time; AI produces the most likely good answer, and most likely means it varies. Ask twice, get two cousins of the same answer. That’s the design, not a fault.
Consistency comes from the library, not the lottery. Keep your best prompts in one doc with [brackets] where the details change, and every reuse starts from your best ask instead of a blank line. When I work with clients we build full prompt guidebooks by role; those libraries are part of what I teach and sell. The method on this page is free, and the method is the valuable bit anyway.
The four mistakes that make outputs generic
- No role. The AI defaults to writing for everyone, which reads as writing for no one.
- No context. It doesn’t know you’re a two-person operation in Roma, not a national chain, unless you say so.
- No format. Ask for an answer and you’ll get an essay. Ask for five dot points and you’ll get five dot points.
- No constraints. Length, tone and what to leave out. The tightest brief wins, same as with any freelancer.
One more, and it’s the quiet one: the words. If your outputs sound like every other AI paragraph on the internet, half the fix is banning the tell-tale vocabulary. I keep a public list of the words to strike; add do not use these words to your prompt and paste the worst offenders in.
Copy-paste starting points
Five to get you moving, distinct from the five jobs in the time-back piece. Swap the brackets, keep the shape.
- Customer enquiry: You are my front-of-house writer. I am a [business type] in [town]. I need a reply to this enquiry that answers their question, offers [next step] and sounds like a human wrote it. Return it under 100 words: [paste enquiry].
- Month of social ideas: You are a content planner for small local businesses. I am a [business type] whose customers care about [two things]. I need 12 post ideas for [month], one line each, no hashtags. Return them as a numbered list grouped by week.
- Job ad: You are a recruiter who writes ads people actually read. I am hiring a [role] for a [size] team in [location]. I need a job ad in a friendly, no-corporate-speak tone covering [three duties] and [pay range]. Return it under 250 words.
- Review response: You are my customer relations writer. I am the owner and I sign every reply personally. I need a response to this [positive or negative] Google review that thanks them, addresses [the specific point] and stays gracious. Return it under 80 words: [paste review].
- Supplier email: You are a commercial negotiator who keeps relationships intact. I am a long-term customer of [supplier]. I need an email asking to revisit [terms] given [reason], firm but friendly. Return it under 150 words with a clear ask in the last line.
A tip from my own working week: talk your prompts instead of typing them. Voice dictation gets the context out of your head faster than your thumbs ever will, and rambling context is still context.
What to do on Monday morning
- Save the four-part template where you can reach it. Notes app, sticky note on the monitor, wherever you’ll actually look.
- Start your library. One doc, one prompt added per week, [brackets] where the details change. In three months you’ll have the most useful document in your business.
- Set up your AI properly once, so it knows who it’s talking to before you ask anything. My Personalise Your LLM guide walks you through it for all four major platforms in about ten minutes each.
Set up your AI properly once.
The Personalise Your LLM guide walks you through it for all four major platforms in about ten minutes each, so your AI knows who it's talking to before you ask anything.
Get the Personalise Your LLM guide →Questions people ask
Do the same prompts work in every AI tool?
The structure does; the results vary. Claude follows detailed instructions closely, ChatGPT is more confident and needs firmer constraints, Gemini can read your Google Workspace files where the others can't see them. Write the prompt once, expect to tweak it per platform. Choosing between the platforms is its own decision.
Why do I get different answers to the same prompt?
AI is a probability engine. It produces the most likely good answer each time, and most likely varies by design. If you need consistency, save your best prompt and reuse it exactly; the tighter the ask, the narrower the range of answers.
How long should a prompt be?
As long as the context needs and no longer. The four-part ask usually lands between 30 and 80 words. Under 10 words you're making the AI guess; past 200 you're probably burying the task. Detail about you always beats instructions about style.
Human-led. AI-leveraged. My philosophy, my business, this article. The Augmented Workforce in action.
Drafted with Ada, my AI collaborator. Reviewed, shaped and signed off by me. How I work with AI· Tracy Sheen CSP