Prompt Engineering: A Practical Guide for Business, With Examples

Most “bad AI results” are actually bad prompts

If you’ve tried one of the big AI tools and come away thinking it’s overhyped, it’s worth a second look. Nine times out of ten the tool isn’t the problem. The instructions are. Modern models are very good at doing what you ask, and not much good at guessing what you meant. Tighten the ask and the output usually changes from “generic waffle” to “I could actually send this.”

That’s all prompt engineering really is: the small craft of asking well. You don’t need to be technical, and you don’t need to memorise anything clever. The techniques below come straight from the published guidance of the people who build these models, and each one comes with a before-and-after you can copy today.

1. Be specific: the model takes you literally

The single biggest improvement most people can make is to stop being vague. A short, fuzzy prompt gets a short, fuzzy answer.

Before: “Write a reply to this customer complaint.”

After: “Write a reply to the customer email below. Apologise for the late delivery, explain that it was a courier delay outside our control, offer a 10% discount on their next order, and keep it warm but professional. Around 120 words. Sign off from ‘The team at [Business]’.”

Same tool, completely different result. You’re not hoping it reads your mind; you’re telling it the tone, the length, the facts and the outcome you want.

2. Give it the context, including the “why”

Models do better work when they understand the situation, just like a new starter would. A line or two of background is usually enough.

Before: “Summarise this report.”

After: “I’m presenting this report to non-technical board members who have ten minutes and care mainly about cost and risk. Summarise it in five short bullet points they could skim before the meeting.”

Telling it who the summary is for and why changes what it chooses to keep.

3. Show it an example (this is the big one)

If you want output in a particular style, show it one. Giving an example or two, sometimes called “few-shot” prompting, is one of the most reliable techniques there is, and it’s still strongly recommended by model makers like Anthropic and OpenAI.

Prompt: “Write product descriptions in the style of these two examples, then write one for the third product.

Example 1 (Oak Dining Table): ‘Solid British oak, built to be argued over at Christmas for the next forty years. Seats six, seats eight if everyone’s friendly.’

Example 2 (Wool Throw): ‘The blanket that ends the thermostat debate. Lambswool, machine-washable, suspiciously hard to give back once someone’s claimed it.’

Now write one for: Ceramic Mug, 350ml, dishwasher safe.”

One good example teaches voice better than a paragraph of adjectives.

4. Give it a role

Telling the model who to be sets the tone and the level of detail without you having to spell it out.

Before: “Give me feedback on this job advert.”

After: “You’re an experienced recruiter who hires for small businesses. Review this job advert and tell me three things that would put off good candidates, and how to fix each one.”

The role does a lot of quiet work: it sets expertise, tone and what “good” looks like.

5. Ask for the format you actually want

If you know what shape the answer should take, say so. It saves you reformatting afterwards.

Before: “Compare these three suppliers.”

After: “Compare these three suppliers in a table with columns for Price, Lead Time, Minimum Order and one Risk to Watch. Then add a single line recommending which to trial first and why.”

You can ask for whatever fits where it’s going next: a table, a list, an email, a script, or a one-line answer.

6. Let it say “I don’t know” and give it the facts

This is the technique that stops the embarrassing mistakes. Left to invent, a model will sometimes fill a gap with a confident guess. Two habits fix most of it: paste in the real source material, and explicitly give it permission to admit uncertainty.

Prompt: “Using only the policy document pasted below, answer the customer’s question. If the answer isn’t in the document, say ‘I’m not certain, please check with us directly’ rather than guessing.

[paste your returns policy]

Customer question: ‘Can I return an item after 40 days if it’s unopened?’”

This is also a gentle introduction to what’s sometimes called context engineering: the idea that giving the model the right information matters as much as giving it the right instructions. For anything where accuracy counts, feeding it your real documents beats relying on its general knowledge every time.

Two habits that beat any single trick

Treat the first answer as a draft. The quickest path to a great result is often a quick second instruction: “make it shorter,” “less salesy,” “add a line about the warranty.” You’re having a conversation, not casting a spell.

Keep your good prompts. When one works, paste it into a simple document. Within a few weeks you’ll have a small library (customer replies, quote drafts, post ideas) that anyone on the team can reuse. That, more than any individual tip, is what turns AI from a novelty into something that quietly saves hours.

Pro Tip: The businesses that get the most from AI aren’t the ones with the cleverest prompts. They’re the ones who picked one repetitive task and got it genuinely reliable before moving on. Our free AI Adoption Playbook walks through how to choose that first task.

Want help putting this to work on a real task?

A free 30-minute call with an engineer. Bring a job that eats hours every week, and we’ll talk through whether AI can take it off your plate.

Book a free call