Technology & AI
What is prompt engineering? A plain-English guide
By the OBA team · 24 July 2026 · 4 min read
Prompt engineering is the skill of writing instructions for AI tools so they produce useful, checkable results. That is the whole definition. There is no code involved and no secret vocabulary: just clear, specific instructions, plus the judgement to check what comes back before you use it.
The name sounds grander than the work. When you type a request into a chat-based AI tool, the quality of what you get back tracks the quality of what you put in. People who write clear briefs get drafts they can actually use. People who type six vague words get something generic, then quietly decide the tool is overrated. The tool was fine. The brief was the problem.
This guide covers why vague prompts fail, the five ingredients of a good one, how to work with the output, and how to tell whether the basics in this post are all you need.
Why vague prompts fail
"Write me a marketing email" is a real prompt people type, and it disappoints every time. Not because the tool is weak, but because that instruction could describe ten thousand different emails. Asked for the average, the tool produces the average: generic, safe and slightly off.
The fix is the same one you would use with a new colleague. You would never hand a new starter a task in five words with no context. You would explain who the email is for, what it needs to achieve, what tone your company uses and what has been tried before. AI tools need exactly that briefing, every single time, because they start each conversation knowing nothing about you, your customers or your standards.
The five ingredients of a good prompt
Most strong prompts contain the same five ingredients. You will not need all five every time, but when an output disappoints, one of them is usually missing.
Role: tell the tool who to be. "You are an experienced UK bookkeeper explaining VAT to a new client" produces a very different answer from no role at all.
Context: the background it cannot guess. Your audience, your product, your constraints, what has already been decided.
Task: one clear instruction. If you need three things, ask for them one at a time or as a numbered list.
Format: say what the output should look like. A table, a bulleted summary, a short paragraph, an email with a subject line.
Examples: show one sample of what good looks like. A single example often improves the result more than a paragraph of description.
Mentioned in this article
Certificate in Prompt Engineering
Self-paced · verifiable certificate · £79
Treat the first answer as a first draft
Put together, the vague email prompt becomes something like: "You are a copywriter for a small UK garden centre. Write a friendly email to past customers announcing our spring workshop programme. Keep it under 150 words, plain English, no hard sell, and end with a link to book." Same tool, completely different outcome.
From there, prompting is a conversation, not a slot machine. The first output is raw material: ask for a shorter version, a different tone, three alternatives to the weakest line. Iterating in follow-ups is normal, and it is usually faster than trying to compose one perfect mega-prompt up front.
Then check the work. AI tools state wrong things with complete confidence, so any fact, figure or name in the output needs verifying before it goes anywhere that matters. You are accountable for what you send, not the tool. The official guides from Anthropic and OpenAI come back to the same core advice in different words: be clear, give context, show examples, and iterate.
Where prompting matters at work
The obvious wins are first drafts: emails, reports, job descriptions, meeting summaries, rough notes turned into an agenda. The pattern that separates useful from gimmicky is simple. Prompting pays off on tasks where you can supply good inputs and check the output yourself: summarising a document you have read, drafting a reply you will edit, restructuring information you already understand.
It pays off least where you cannot check the result. If a subject is completely new to you, you will not spot the confident error in paragraph three. Build enough understanding to review the output before you rely on it, and treat anything you cannot verify as a lead to chase rather than an answer to forward.
A durable skill, not magic words
Sceptics say prompt engineering will fade as models get better at guessing intent. The tools are improving, but the bottleneck was never really the software. It is that most instructions, to humans or machines, are vague. Prompting is a skill like searching was a skill in 2005: invisible to the people who have it, baffling to the people who do not.
That is why the durable part is not a stack of memorised trick phrases. It is clear thinking written down: knowing what you actually want, which context matters and what a good result looks like. Models will keep changing. That will not.
Do you need a course for this?
Honestly, maybe not. If you use an AI tool once a month for a quick draft, the ideas in this post are enough. Save the five ingredients somewhere handy, skim the official docs, and carry on with your day.
The Certificate in Prompt Engineering is for people who want prompting as a working skill: something you use daily, on real tasks, with confidence in the results. The lessons are short and interactive, so you practise on realistic work scenarios rather than watching someone else type. And if your real question is bigger than prompts (where AI fits across your whole role), the Certificate in AI at Work is the better starting point.
Not sure which fits? The free info pack sets out what each course covers, and it costs nothing to find out.
Good to know
Common questions
A small number of dedicated prompt engineering roles exist, mostly inside AI-focused companies. For most people, though, it works like spreadsheet skills: a capability used inside an existing job rather than a job title of its own. The value is in doing your current work faster and better.
No. Prompts are written in plain English, and clear writing matters far more than technical knowledge. Code only enters the picture if you are building software on top of AI models, which is a different discipline entirely.
The trick phrases will date, but the core of the skill will not. Describing what you want clearly, supplying the right context and checking what comes back are durable habits. Tools are getting better at guessing intent, and clear instructions still beat vague ones every time.
Five things: a role for the tool, context it cannot guess, one clear task, the format you want, and an example of what good looks like. Then treat the first answer as a draft and refine it in follow-ups rather than expecting perfection in one go.
Whichever one your workplace allows you to use. The technique transfers between tools, and the official guides from the major providers give near-identical advice, so nothing you learn is wasted if you switch later.