Definition
Prompt engineering is the practice of designing, testing and refining the instructions you give to a generative AI model so that it returns output that is accurate, relevant and consistent. A prompt is everything the model sees before it answers: the question, the role you assign it, the background information, examples, constraints and the format you want back. Prompt engineering treats that text as something you can deliberately design rather than something you type on impulse.
The term started in research labs, where engineers discovered that small wording changes could swing a model's accuracy on benchmark tasks by large margins. Today it is an everyday workplace skill. Anyone who uses ChatGPT, Claude, Gemini or Microsoft Copilot to draft an email, summarise a report or analyse a spreadsheet is doing prompt engineering, whether they call it that or not. The difference between a casual user and a skilled one is that the skilled user knows which ingredients reliably improve results and applies them on purpose.
Why prompt engineering matters
Large language models do not read your mind. They predict the most likely continuation of the text they are given. When a prompt is vague, the model fills the gaps with the most generic answer it can find, which is usually bland, too long or wrong for your context. When the prompt supplies a clear goal, the audience, the relevant facts and the expected format, the space of likely answers shrinks dramatically and the output becomes useful on the first or second try.
Research supports this. In a preregistered Harvard Business School and Boston Consulting Group field experiment with 758 BCG consultants, people using GPT-4 on 18 realistic tasks inside the model's capability completed 12.2% more tasks and worked 25.1% faster, with significantly higher quality. On a task deliberately chosen to be outside the model's capability, consultants using AI were 19% less likely to reach the correct answer. Good prompt engineering is partly about getting better output and partly about knowing when to check, constrain or not use the model at all.
The core ingredients of a good prompt
Most professional prompt frameworks share the same building blocks:
- Context — who you are, what the situation is and what the model needs to know that it cannot guess.
- Role — the perspective the model should adopt, for example "an experienced HR business partner" or "a sceptical financial analyst".
- Action — the specific task, stated with a clear verb: summarise, compare, rewrite, draft, critique.
- Format — the shape of the answer: a table, five bullet points, a 120-word email, JSON.
- Tone and constraints — reading level, length, words to avoid, facts that must not be invented.
Advanced techniques build on these basics. Few-shot prompting adds worked examples. Chain-of-thought prompting asks the model to reason step by step. Iterative prompting treats the first answer as a draft and asks for targeted revisions. Prompt templates turn a prompt that works into a reusable asset the whole team can share.
Example
Imagine a team lead who needs to announce a new hybrid-work policy.
Weak prompt: "Write an email about the new hybrid policy."
The model returns a generic, upbeat email that invents details, ignores the team's concerns and runs to 400 words.
Engineered prompt: "You are an internal communications specialist. Context: our 40-person finance team moves from five office days to three (Tuesday to Thursday) starting 1 March. Many people have raised concerns about desk availability. Task: draft an email from me, the team lead, announcing the change. Format: under 180 words, a short subject line, three bullet points covering what changes, what stays the same and where to ask questions. Tone: warm and direct, no corporate jargon. Do not invent any details that are not listed here."
The second version produces a usable draft in one pass. It is shorter, accurate, addresses the desk concern and needs only light editing. The time saved is small for one email, but across dozens of tasks a week it adds up to hours.
Common mistakes
- Asking for everything at once. Complex tasks work better as a sequence of smaller prompts.
- Leaving out the audience. "Explain this" gives a very different answer from "explain this to a new employee with no finance background".
- Not specifying format. Without a format, models default to long, padded prose.
- Pasting sensitive data. Personal data, client information and confidential numbers should only go into approved, enterprise-grade tools.
- Accepting the first answer. Treat output as a draft. Ask the model to shorten, check assumptions or list what it is unsure about.
Prompt engineering frameworks
Frameworks give people a memorable checklist. CRAFT (Context, Role, Action, Format, Tone) is designed for general workplace use. RTF (Role, Task, Format) is a lightweight version for quick requests. RACE, CO-STAR and PARTS add elements such as audience, style and examples. None is magic; they all remind you to supply the information a model needs. Teams that standardise on one framework find it easier to share prompts, review each other's work and build a library of templates that new staff can adopt from day one.
Is prompt engineering still relevant as models improve?
Newer models are better at guessing intent, so some tricks that mattered in 2023 now matter less. The fundamentals have not changed: a model cannot know your company's context, your audience or your quality bar unless you tell it. As AI moves into agents and automated workflows, prompts become instructions that run thousands of times, which makes clear, tested wording more important, not less. Prompt engineering is evolving into a broader skill sometimes called context engineering — deciding what information, tools and examples a model should have for each step of a task.
Key takeaways
- A prompt is a design artefact: goal, context, role, format and constraints.
- Clear prompts reduce errors, shorten editing time and make output repeatable.
- Frameworks like CRAFT turn good habits into a checklist the whole team can use.
- Always review output and keep sensitive data out of unapproved tools.