Prompt Engineering and In-Context Learning: How to Get Better Results from AI

Ask ChatGPT the same question two different ways and you’ll get two different answers. Sometimes one is brilliant. Sometimes one is rubbish. The gap between them is prompt engineering — and it’s become one of the most useful skills in Britain’s AI toolkit for 2026. UK businesses spent an estimated £4.2 billion on AI tools last year, according to techUK figures, yet most staff still type prompts the same way they’d type a Google search. That’s the problem. Large language models don’t search. They predict. And predicting well needs a different kind of instruction.

What Is Prompt Engineering?

Prompt engineering means writing instructions that steer an AI model toward the answer you actually want. It sounds simple. It rarely is.

When I looked into this properly, the pattern became obvious fast. Vague prompts get vague answers. “Write me a blog post about pensions” produces generic filler. “Write a 600-word blog post explaining UK auto-enrolment pension rules for a 28-year-old first-time employee, in a friendly but factual tone, with three practical examples” produces something usable.

The model hasn’t changed between those two prompts. Only the instruction has. That’s the entire discipline in one sentence. Everything else — examples, structure, reasoning steps — builds on that single idea: specificity beats cleverness.

It also explains why two people using the exact same AI subscription can get wildly different value from it. The tool is identical. The instructions aren’t.

In-Context Learning: Teaching AI Without Retraining It

In-context learning is the technical term for something users do instinctively: teaching the model inside the conversation itself, rather than retraining it.

Traditional machine learning needs new training data and weeks of computation to learn a new task. Large language models like GPT-5 and Claude Opus 4.5 skip that step entirely. Show the model two or three examples of what you want inside the prompt, and it adapts its output to match — instantly, with zero retraining.

This matters for UK small businesses especially. A five-person marketing agency can’t afford to fine-tune a custom model. But they can paste three examples of their brand voice into a prompt and get consistent output within seconds. Cost: effectively nothing. Barrier to entry: gone.

UK investors keep asking about this because it changes the economics of adopting AI. You no longer need a data science team. You need someone who understands how to structure an instruction.

Researchers still argue over exactly why this works. The leading theory is that huge models trained on enough text implicitly learn a kind of pattern-matching machinery during training, which the prompt then activates on demand. Nobody fully agrees on the mechanism. Everybody agrees it’s useful.

Zero-Shot, One-Shot and Few-Shot Prompting

Researchers split in-context learning into three flavours.

Zero-shot means giving the model a task with no examples at all — just an instruction. “Translate this into French.” Works fine for simple, well-known tasks.

One-shot gives a single example first. Useful when format matters — showing the model exactly how you want a table structured, say.

Few-shot gives three to five examples. This is where accuracy jumps hardest. Studies from Anthropic and OpenAI both show few-shot prompts cutting error rates by 20 to 40 percent on structured tasks compared with zero-shot versions.

Short version: more examples, better output, up to a point. Beyond five or six examples, returns flatten out fast. Some models even get confused by too many examples, latching onto irrelevant patterns in the extras rather than the actual task.

Chain-of-Thought Prompting: Getting AI to Show Its Working

Ask an AI to solve a maths problem straight and it sometimes guesses wrong, confidently. Ask it to “think step by step” first, and accuracy climbs sharply.

This is chain-of-thought prompting. It forces the model to lay out reasoning before landing on an answer, the same way a maths teacher demands you “show your working” rather than just writing the final number.

Google’s own research on this technique found accuracy gains of over 30 percentage points on multi-step reasoning tasks when chain-of-thought was added versus a bare prompt. That’s not a rounding error. That’s the difference between a tool you can trust and one you can’t.

It works because language models generate text one token at a time. Forcing intermediate steps gives the model somewhere to “think out loud,” and each step constrains the next, cutting down on wild leaps to a wrong conclusion.

Newer “reasoning” models bake this in automatically, thinking through a problem internally before answering. Older or smaller models still need it spelled out explicitly. Worth checking which type you’re using before assuming the trick isn’t needed.

System Prompts vs User Prompts

Most people only ever write user prompts — the message typed into the chat box. But every serious AI deployment also uses a system prompt, sitting invisibly underneath.

The system prompt sets the model’s role, tone and boundaries before the conversation even starts. “You are a UK tax adviser. Never give investment advice. Always cite HMRC guidance where relevant.” That instruction shapes every single reply that follows.

Get the system prompt wrong and no amount of clever user prompting fixes it. Get it right and even a lazy user prompt produces something usable. It’s the foundation, not the finishing touch.