What Is Prompt Engineering? A Plain-English Guide with Examples

Prompt engineering is the practice of writing instructions that get a reliable, usable result from an AI model — first time, and again next week. It is not a set of magic words. It is ordinary professional communication, written down properly.

If you have ever asked a model for "a summary of this report" and received something bland, and then asked a colleague the same way and received something useful, the difference was not intelligence. Your colleague already knew the audience, the deadline, the house style and what you were going to do with the summary. The model knows none of that unless you say it.

Why the same question gives different answers

A language model predicts the most plausible continuation of your text. Vague input has many plausible continuations, so you get the average of them — generic, hedged, mid-length. Specific input narrows the space of plausible answers, so the output lands closer to what you wanted.

That single mechanism explains almost every prompting tip you will read. Add context, and you rule out irrelevant answers. Add a role, and you shift the vocabulary and level of detail. Add a format, and you remove the guesswork about structure. Nothing mystical is happening.

The CRAFT structure

Instead of collecting tricks, use one repeatable structure. We teach CRAFT because it covers the five things a model is actually missing:

• Context — the situation, the audience, the background facts, the constraints
• Role — the perspective it should write from, and the expertise level of the reader
• Action — the single, concrete task, stated as a verb
• Format — length, structure, sections, tone of output
• Tone — formal, plain, warm, blunt; and what to avoid

You will not need all five every time. You will need Context and Action every time.

A worked example

Weak prompt: "Write an email about the project delay."

CRAFT prompt: "Context: our website redesign will miss the 30 June launch by about three weeks because the content migration is slower than planned. The recipient is the client's marketing director; she has already told her board the June date. Role: you are the account lead, and you own the miss. Action: draft the email that tells her about the delay and proposes a new date of 21 July, plus a partial launch of the homepage on 30 June. Format: under 180 words, no bullet points, one clear ask at the end. Tone: direct and accountable, no corporate softening, do not use the word 'unfortunately'."

The second prompt takes forty seconds longer to write and saves three rounds of editing. That trade is the whole discipline.

Few-shot examples: the fastest quality upgrade

When the output has a house style — support replies, product descriptions, job ads — paste two or three examples of good past work before your instruction and say "match the style of these". The model infers rules you would struggle to describe: sentence length, how you open, how much hedging you tolerate, which words you never use.

Two strong examples beat a paragraph of style adjectives almost every time.

Iteration beats one-shot magic

Professionals rarely get the final version on the first prompt, and they do not start over. They correct:

• "Second paragraph is too vague — name the specific risk."
• "Keep the structure, halve the length."
• "You invented a statistic. Remove any number I did not give you."

Each correction is information the model did not have. Treat the conversation as editing, not as gambling on a better first prompt.

What prompt engineering does not fix

It does not make the model truthful. Names, figures, dates, citations, legal and medical claims still need verification against a source you can point to. It does not make private data safe to paste into a consumer tool. And it does not replace your judgment about whether the output is any good — you remain the person whose name is on the work.

Turn what works into templates

The real return arrives when a prompt stops being disposable. When a prompt produces a genuinely good result, save it with the variable parts marked, for example [audience], [deadline], [key numbers]. A dozen of those, organised by task, is a personal prompt library — and it compounds, because each use is faster than the last.

Where to learn this properly

Reading about prompting is not the same as being able to do it under time pressure on your own work. If you want structured practice, Module 1 of AI POWERED is free, Module 2 covers CRAFT, context, few-shot prompting, iteration and templates in seven lessons, and the prompt library gives you starting points by role. Each module ends with an assessment you have to pass at 80% before moving on, so you find out whether it stuck.

Frequently asked questions

What is prompt engineering? Writing AI instructions that reliably produce a usable result — supplying the context, role, task, format and tone the model would otherwise guess.

Do I need to be technical? No. It needs clear writing and knowledge of your own work, not programming.

What is the CRAFT framework? Context, Role, Action, Format, Tone. The full breakdown is in the CRAFT prompt framework.

How long does it take to get good? Most professionals see a clear jump within a few hours of structured practice, and reach reliable results once they keep a small template library.

Is it safe to use these prompts on company data? Only with rules. Run every task through the SAFE framework before pasting anything sensitive.

Related reading: The CRAFT prompt framework for the structure itself, the SAFE framework for using AI on real work data, and What is Titan OS? for an example of reading a technology decision clearly.

Also useful: How to use AI at work and Is ChatGPT safe to use at work?.

Keep reading

• Prompt engineering frameworks compared: CRAFT, RTF, RACE, CO-STAR, PARTS