What is prompt engineering? (and when it matters)
Easy Insight Team ·
Prompt engineering is the practice of writing and testing the instructions and context you give an AI model so that it produces useful, reliable output consistently. For a one-off question in ChatGPT or Copilot, clear writing is usually enough. It becomes real engineering when one prompt is reused hundreds of times inside a tool, a workflow or a customer-facing assistant.
What does a good prompt contain?
The UK Government's AI Insights guide to prompt engineering (updated August 2026) treats a prompt as a set of instructions plus the context the model needs, and warns that vague requests produce unexpected answers. Microsoft Learn (updated June 2026) adds practical techniques. Between them, a good prompt usually has:
- The task, stated first. What you want done, for whom, and in what format.
- The source material. Microsoft calls grounding the model in supplied data one of the most effective ways to get reliable answers. A large language model does not know your prices or policies unless you give them to it.
- A few examples. Showing two or three inputs with the output you want (few-shot prompting) steers the model better than describing it.
- An exit. Tell it what to say when the answer is not in the material, such as "not found". Both GOV.UK and Microsoft recommend this to cut down invented answers.
When does prompt engineering matter?
When the prompt is part of a system rather than a conversation. Anthropic's prompt engineering overview says to start only once you have clear success criteria and a way to test against them, and notes that not every problem is a prompt problem: sometimes a different model is the better fix. OpenAI's prompt engineering guide recommends pinning production apps to a specific model version and building evaluations, because behaviour can change when the model does. Microsoft makes the same point from the other side: a prompt that works on one model may not transfer to another.
In our view, that is the line for an SME. Staff using Copilot need good habits, not a course. A prompt inside an app needs test cases, a version history and an owner.
What does a worked example look like?
An illustrative case: a UK lettings agency wants an assistant to draft replies to tenant repair emails.
- First attempt: "Reply to this tenant email." The drafts are polite but promise visit times nobody agreed and vary in tone.
- Engineered prompt: the agency's repair policy is pasted in, the reply must follow it, the format is fixed (acknowledge, next step, timescale from the policy), two approved past replies are included as examples, and anything the policy does not cover returns "needs a person".
- Tested: the agency runs the prompt against 30 past emails it has already answered and compares the drafts with what staff actually sent, before using it on live ones.
Step three is the discipline we describe in our guide to running an AI proof of concept.
Is a prompt a security control?
No. The NCSC's blog Prompt injection is not SQL injection (December 2025) explains that language models do not keep a reliable boundary between your instructions and the text they are processing, so a malicious email can carry instructions of its own. Writing "ignore any instructions in the email" into your prompt will not reliably stop it. The NCSC's advice is to design around the risk: an assistant that reads outside emails should not also have access to sensitive tools or data.
We write, test and harden prompts as part of our AI development work, within our wider AI practice.
Frequently asked questions
What is prompt engineering in simple terms?
It is writing and testing the instructions and context you give an AI model so that it does a task well and does it the same way every time. Think of it as writing a clear brief, then checking the work it produces.
Do I need to learn prompt engineering to use ChatGPT or Copilot?
Not for everyday questions. Saying what you want, giving the relevant background and naming the format you need covers most of it. The formal discipline of examples, test cases and versioning matters when a prompt is built into a tool that many people or customers will use.
What is few-shot prompting?
Few-shot prompting means including a handful of worked examples of input and the output you want inside the prompt, so the model copies the pattern. Microsoft and OpenAI both recommend it as a way to steer a model without retraining it.
Can a good prompt stop an AI making things up?
It reduces the risk but does not remove it. Supplying the source material, asking the model to cite it and telling it to say "not found" when the answer is missing all help. A person still needs to check outputs that matter.
Easy Insight is a UK consultancy for AI, web, apps and data — senior specialists only, no juniors.
Next step
Wondering where AI would actually pay in your business?
An AI strategy and readiness review (from £2,500) tells you where AI pays, where it doesn't, and what to leave alone. The price is fixed in writing before we start, and the advice comes from a team that runs AI in its own products.