The way you ask makes a huge difference to what you get back. A good prompt is like a good brief to a new member of staff: clear about the goal, with the background they need and a picture of what "good" looks like. That's all "prompt engineering" really is.
Vague prompt
Prompt: "Write something about AI."
Problem: No goal, audience, length or context - you'll get something generic.
Specific prompt
Prompt: "Write a 500-word blog post for UK small business owners explaining how an AI chatbot can improve customer service, with 3 practical benefits and one honest limitation."
Why it's better: Clear topic, audience, length and structure.
Seven techniques that work
Break your request into clear parts. This is the single biggest improvement most people can make.
Goal: Write a reply to a customer complaint.
Context: We are a family-run bakery. The customer's birthday cake was delivered an hour late. Their email is pasted below.
Constraints: Apologise sincerely, no excuses, offer a 20% refund. Under 150 words.
Format: Ready-to-send email with a subject line.
Attach the PDF, spreadsheet, screenshot or email thread rather than summarising it yourself. All the main AI tools can read files, and the answer will be based on the real thing.
"Attached are our last 12 months of sales by product. Which three products are growing fastest, which are declining, and what might explain it? Show the figures you used."
Tell the AI who it should be and who it is writing for.
"You are an experienced small business marketing consultant. A local bakery wants to increase online orders. Suggest 5 specific actions they can take this month with a budget under £500. The owner is not technical, so avoid jargon."
Give one or two examples of the style or format you like (sometimes called "few-shot prompting").
"Write customer service replies in this style:
Customer: 'My order is late'
Reply: 'I'm sorry your order hasn't arrived yet. I've checked and it's with the courier for delivery tomorrow - here's the tracking link.'
Now write a reply to: 'The product doesn't work as advertised'"
For bigger tasks, let the AI find out what it needs before it starts.
"I need a staff handbook section on hybrid working. Before you write anything, ask me up to 10 questions about how our business works."
Modern "reasoning" models work through problems step by step on their own. For complex decisions choose the thinking or reasoning mode, and ask it to plan first and explain its reasoning.
"We have £10,000 to spend on AI. We get 50 customer enquiries a day and spend 2 hours daily on customer service emails. Plan out the options, compare costs and benefits, then recommend one."
Instead of re-explaining your business every time, save your background information, tone of voice and templates in a Project (ChatGPT or Claude), a Custom GPT, a Gemini Gem or Copilot agent. Everyone on your team then starts from the same, well-briefed AI.
A "Marketing Assistant" project containing your brand guidelines, product list, three of your best past emails and the instruction: "Always write in British English, friendly but professional, never use exclamation marks."
Multimodal prompting - images, files and voice
ChatGPT, Claude, Gemini and Copilot can all work with images, screenshots and documents, and most can take voice input too.
Prompt (with a photo attached): "Here's a photo of our restaurant's menu board. Suggest 3 changes to make it easier to read and increase the average order value."
Why it works: The AI sees exactly what your customers see, and the task is specific.
Drag the slider (or click the image) to compare the menu board before and after.
What the AI suggested
- Simplify the layout: Simplify the layout so the sections are easier to scan and there is less decorative clutter.
- Align the prices: Align prices consistently so customers can compare items quickly.
- Highlight higher-value items: Highlight profitable specials, drinks, desserts and add-ons to gently encourage higher-value orders.
All three suggestions have been applied in the revised board - a quick redesign that took minutes to plan with AI.
Always check the output
Check facts and figures
AI can still "hallucinate" - state things confidently that are wrong. Ask for sources and double-check numbers, names, dates and legal points.
Iterate
Treat the first answer as a draft. Say what to change - "shorter", "more formal", "use UK spelling" - and it will improve.
Mind the data
Only use business accounts for work, and don't paste in personal data you don't need to. See Sensitive Data and AI.
Last reviewed: 23-Sep-2026