How to write a prompt that actually works at work

Most people who struggle with AI tools at work are not using the wrong tool. They are writing the wrong prompt. The output looks vague, slightly off, or confidently wrong, and the natural conclusion is that the technology is not ready. In most cases, the technology is ready. The instruction was not.

This article is about the method behind writing prompts at work, not a list of tricks that will be obsolete when the next model lands. The principles here come from the same framework that underpins Unit 02 of the NCFE Level 2 Certificate in Artificial Intelligence (AI) for the Workplace: understanding what an AI system needs in order to produce useful output, and developing the critical judgement to evaluate what comes back. Those two skills matter far more than memorising particular prompt formats.

Why most prompts fail

When you open a meeting, you do not walk in, say one sentence, and expect a finished report on your desk. You share background, context, constraints, audience and purpose. A language model needs exactly the same information, and unlike a colleague, it cannot ask a follow-up question in the corridor. It can only work with what you give it in the moment.

The single most common failure in writing prompts at work is treating the prompt as a search query. Typing "write a report on our Q3 performance" is the equivalent of walking up to a new contractor and saying "do the work." The contractor needs to know what sector you are in, who will read the report, what decisions it needs to support, how long it should be, and what tone is appropriate for your organisation. So does the model.

The four things every useful prompt must include

Experienced practitioners tend to converge on a small number of structural elements, regardless of which model they are using. These are not rigid fields to fill in order. They are categories of information that prevent the model from guessing, and guessing badly.

1. Role and context

Tell the model who it is working as and, more importantly, who you are and what your situation is. These are different things. "You are a professional editor" sets the model's behaviour. "I am a compliance officer at a mid-sized UK financial services firm and I need to communicate a process change to colleagues who are not technical" sets your context. Both matter.

Context the model cannot see includes your industry, your organisation's tone of voice, the seniority of the audience, and whether the output will be read on screen or spoken aloud in a meeting. None of that is visible to the model unless you say it.

2. The task, stated precisely

Be specific about the output you want, not just the subject. There is a large difference between "write something about the new data protection policy" and "draft a 200-word internal announcement, suitable for an all-staff email, explaining that we are updating our data retention schedule from 1 September, why it matters to staff, and what they need to do before that date." The second prompt eliminates a dozen guesses the model would otherwise have to make.

If you are asking for how to write a good prompt, the answer starts here: the task description is the most important single element. Everything else provides context around it.

3. Constraints and format

Specify length, structure, tone, what to include and, critically, what to leave out. If you do not want the output to make specific claims about competitors, say so. If it must avoid jargon because the reader is not technical, say so. If it needs to end with a single call to action, describe that call to action.

Constraints are not limitations on the model's creativity. They are the professional brief, and a good brief produces better work from any writer, human or otherwise.

4. The standard you will use to judge it

This is the element most people omit entirely. Tell the model what good looks like. "The output should be clear enough that a new employee on their first week could act on it without guidance" is a concrete standard. "It should sound professional" is not, because professional means something different in every organisation.

The model cannot read your mind, your brand guidelines, or your organisation's history. Every piece of information it is missing is a gap it will fill with a plausible guess. Your job when writing prompts at work is to close those gaps before you hit send, not after you read the output.

A worked example: the meeting summary

Here is an ordinary office task. You have just come out of a 45-minute project meeting and you need a summary to send to three stakeholders who were not present. One is a director who wants the headline decisions only. One is a project manager who needs actions and owners. One is a supplier contact who needs to know their deliverables and the revised timeline.

A weak prompt: "Summarise this meeting."

A prompt that works:

"I have pasted rough notes from a 45-minute project meeting below. Please produce three short summaries from the same notes, each no more than 150 words. The first is for a director: lead with the two decisions made and nothing else. The second is for the internal project manager: a bulleted action list with owner names and deadlines as they appear in the notes. The third is for an external supplier: only what relates to their deliverables and our revised timeline for sign-off. Do not add information that is not in the notes. If something is unclear in the notes, flag it with [CHECK] rather than guessing."

That second prompt takes about 90 seconds to write. The output requires minimal editing. The first prompt produces something generic that you will spend longer correcting than writing from scratch.

The instruction the model cannot see: your organisation's context

Language models are trained on large bodies of public text. They have never read your internal style guide, your sector's regulatory requirements, your organisation's values statement, or the three previous emails on this thread. If any of those things matter to the output, you need to supply them or summarise them in the prompt.

For regulated sectors this is particularly important. A prompt that asks for "a client-facing explanation of how our investment decisions are made" will produce something plausible and potentially problematic. Adding "this must not constitute financial advice, must be factually accurate to the following process description, and must include a statement directing the reader to seek independent advice" changes the output and, more importantly, changes your accountability for it.

Speaking of accountability: if you are thinking carefully about where responsibility sits when AI output is used in a professional context, the article Who is accountable when an AI system gets it wrong? on this blog addresses that question directly. It is worth reading before your output gets closer to a client.

Judging what comes back

The second skill, and the one that separates a professional who uses AI well from one who merely uses AI, is critical evaluation of the output. Pasting the model's response directly into your work without reading it carefully is not a time-saving strategy. It is a risk transfer, and not in your favour.

When you read the output, ask four questions.

• Is every factual claim in here something I can verify, or is the model presenting a confident-sounding guess as established fact?
• Does this reflect the actual context I provided, or has the model drifted back to generic language?
• Would my intended reader understand this, or has the model written for a different audience than the one I specified?
• If I put my name on this, would I be comfortable defending every sentence?

If the answer to any of those is no, you iterate. You go back to the prompt, identify which piece of context was missing or ambiguous, and try again. This is not a failure of the technology. It is the normal working process, and it gets faster with practice.

AI prompts for work: a comparison of prompt quality

Task	Weak prompt	Stronger prompt	What the stronger prompt adds
Internal announcement	Write about the new policy.	Draft a 150-word all-staff email announcing our updated expenses policy, effective 1 August. Audience: UK office staff, not technical. Tone: clear and friendly. Include one action and one deadline.	Audience, length, tone, required content
Meeting summary	Summarise this meeting.	From the notes below, produce a 100-word summary for the project manager only, listing actions, owners and deadlines. Flag anything unclear as [CHECK]. Do not include discussion that did not result in a decision or action.	Audience, format, scope, quality flag
Customer email response	Reply to this complaint.	Draft a response to the complaint below. Tone: empathetic but professional. Do not admit liability. Acknowledge the issue, explain the next step, and give a realistic timeframe. Max 120 words.	Tone, constraint, structure, length
Briefing document	Write a briefing on hybrid working.	Write a two-page briefing for our senior leadership team on hybrid working policy options. UK context. Include three options with pros and cons for each. End with a recommendation section but do not make the final recommendation for us.	Audience, scope, structure, boundary

Why method matters more than model

The specific phrasing that produces a good result from one model may not transfer perfectly to the next version, or to a different tool. What does transfer is the underlying discipline: define the audience, state the task precisely, provide constraints, supply the context the model cannot access, and evaluate what comes back against the brief you gave it.

This is also why formal learning in this area has practical value. The NCFE Level 2 Certificate in Artificial Intelligence (AI) for the Workplace covers not just how to construct prompts but how to evaluate AI outputs critically and how to apply AI tools responsibly within a workplace setting. Those are transferable professional skills, not platform-specific tricks.

If you want to understand what a structured programme in this area looks like and whether it fits around your working week, this article on getting a qualification in using AI at work explains the options, the entry points and what regulated qualifications actually require of learners.

A note on iteration

Professional writers do not produce a finished article in one draft. Professional designers do not present the first sketch. The same is true of working with AI. The first output is a draft to react to, not a finished product. Building iteration into your process, and treating it as normal rather than as evidence that something has gone wrong, will improve your results more reliably than any single prompt technique.

Keep a personal log of prompts that worked well for recurring tasks. An email you draft well once can become a reusable template. The investment in writing a careful prompt the first time pays back every time you reuse it.

Getting started this week

Pick one task you do regularly that currently takes more time than it should. A weekly report, a standard client update, a first draft of a proposal. Write the prompt properly using the four elements above. Evaluate what comes back honestly. Iterate once. Compare the total time spent to your previous approach.

That single experiment, done carefully and critically, will teach you more about how to write a good prompt than reading any number of prompt libraries.