Every time you type a question into ChatGPT, Gemini or Claude, you’re writing a prompt. It’s the most ordinary part of using AI — and also the part that makes the biggest difference to whether you get something useful back or something generic and vague.
I see the same pattern again and again: someone tries an AI tool, types a one-line request, gets a bland answer and decides the tool “isn’t very good”. Very often the tool was fine. The prompt just didn’t give it enough to work with. In this post I’ll explain what a prompt is, what a good one contains, and a few simple techniques that reliably improve results.
A prompt is your instructions to the model
A prompt is the text you give an AI model to tell it what you want. It can be a question, an instruction, a document to work on, examples of what good looks like — or all of these together. The model reads the prompt (as tokens, which we covered in What are tokens?) and generates a response based on it.
OpenAI describes prompt engineering as “the process of writing effective instructions for a model, such that it consistently generates content that meets your requirements”[4]. I like that definition because of the word consistently. A prompt that works once is luck. A prompt that works every time is a tool you can build a business process on.
The key mindset shift: treat the model like a very capable new colleague who knows nothing about your business, your customers or what you’re trying to achieve. It can only use what you tell it.
The five ingredients of a good prompt
The major AI providers all publish prompting guides, and they agree on the fundamentals: be clear and specific about what you want, give relevant context, show examples, and say what format the answer should take[3][4][5]. I group that advice into five ingredients:
| Ingredient | What it answers | Example |
|---|---|---|
| Role | Who should the model act as? | “You are an experienced customer service lead.” |
| Task | What exactly needs doing? | “Draft a reply to this complaint.” |
| Context | What background does it need? | “The order was late due to a supplier delay. We offer free shipping on the next order.” |
| Format | What should the output look like? | “Under 120 words, friendly but professional, Australian spelling.” |
| Examples | What does good look like? | “Here are two replies our team wrote that customers loved: …” |
You don’t need all five every time. A quick question needs none of them. But when the answer matters — anything customer-facing, anything you’ll reuse — working through the list makes a noticeable difference.
Before and after
Here’s what that looks like in practice.
Before: “Write a reply to this customer complaint.”
After: “You are a customer service lead at a small Australian homewares store. Draft a reply to the complaint below. The order was delayed by a supplier issue, which is now fixed. Apologise once, explain briefly, and offer free shipping on their next order. Keep it under 120 words, warm but professional, using Australian spelling. Don’t promise a delivery date.”
The first prompt forces the model to guess the tone, the reason for the delay, what you’re willing to offer and how long the reply should be. The second leaves almost nothing to guess. Same model, very different result.
Zero-shot, few-shot and showing your working
You’ll hear a bit of jargon around prompting. Here are the three terms worth knowing.
Zero-shot
You simply ask, with no examples. “Summarise this article in three bullet points.” This works well for common, clear-cut tasks.
Few-shot
You include a handful of examples of the input and the output you want, then give the real input. The research paper that introduced GPT-3 showed that large language models can pick up a new task from just a few examples in the prompt, without any retraining[1]. Google’s Gemini guide goes as far as recommending that you always include few-shot examples in your prompts[5]. In my experience examples are the single most effective way to get a consistent tone and format, because showing beats describing.
Chain of thought
For problems that need reasoning — maths, logic, multi-step decisions — it helps to have the model work through the steps before giving an answer. Researchers at Google found that including examples with step-by-step reasoning (“chain-of-thought prompting”) significantly improved large models’ performance on arithmetic, commonsense and symbolic reasoning tasks[2]. Many newer models now do some of this thinking automatically, but asking for the reasoning can still help, and it makes the answer easier for you to check.
System prompts: instructions that stay in place
If you use AI through a business tool or build it into your own product, you’ll come across the system prompt (sometimes called developer instructions). This is a standing set of instructions that sits behind every conversation — who the assistant is, what it should and shouldn’t do, what tone to use. OpenAI’s documentation explains that these high-level instructions take priority over what the user types in an individual message[4].
For a business, the system prompt is where your policies live: “Never give legal advice”, “Always refer refunds over $200 to a human”, “Use our brand voice”. It’s worth treating it like any other important business document — written carefully, reviewed and version-controlled.
Anthropic’s guidance suggests that before you spend time on prompt engineering, you should have a clear definition of success and a way to test against it[3]. In plain terms: collect 10–20 real examples, decide what a good answer looks like, and check every prompt change against that set. Otherwise you’re just tweaking by feel.
What I want you to take away
- A prompt is your instructions to the model. The model only knows what you tell it, so vague prompts get vague answers.
- Use the five ingredients: role, task, context, format and examples — especially for anything you’ll reuse.
- Examples are your strongest lever. Few-shot prompting is the quickest route to consistent tone and format.
- Ask for the reasoning on hard problems. Step-by-step thinking improves results and makes errors easier to spot.
- Treat system prompts as policy. Write them carefully and test changes against real examples before rolling them out.
Working out where AI fits in your business?
I help Australian teams cut through the hype and pick AI projects that actually pay off. If you'd like a second opinion on an idea, I'm happy to talk it through.
Get in touch →References
- Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. NeurIPS 33. https://arxiv.org/abs/2005.14165
- Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 35. https://arxiv.org/abs/2201.11903
- Anthropic. Prompt engineering overview. Claude Developer Platform documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview
- OpenAI. Prompt engineering. OpenAI API documentation. https://developers.openai.com/api/docs/guides/prompt-engineering
- Google. Prompt design strategies. Gemini API documentation. https://ai.google.dev/gemini-api/docs/prompting-strategies