Writing effective prompts

Ask most people what went wrong when an AI tool disappointed them, and you’ll hear about the tool. The output was generic, they say. It missed the point. It sounded like a press release written by a committee. All true — and almost none of it the model’s fault. The same model that produced the wallpaper also produces precise, usable drafts for millions of professionals every day, and the difference between the two outcomes sits almost entirely in the sentence the human typed first. This guide is about that sentence: what it needs to contain, what it needs to omit, and the small craft habits that turn an AI tool from a slot machine into a colleague.

Why ‘write something about X’ fails

Start with empathy for the failure. When you type a vague request, the model must guess every unspecified dimension: who this is for, what it’s for, how long it should be, what tone fits, what to include, what to avoid, and what success looks like. It resolves each guess toward the statistical middle — the average of everything it has read. The statistical middle of business writing is where personality goes to die. So the generic output isn’t the machine failing to understand you; it’s the machine understanding you perfectly and correctly reporting that you gave it nothing to work with. The fix isn’t a better model. It’s a better brief.

This reframing matters practically, because it moves the work to where your leverage is. You cannot make the model smarter, but you can make your request smarter in about ninety seconds — and that ninety seconds is the highest-return writing investment most professionals will ever make. An email brief that produces a sendable draft beats five minutes of prompt-gambling followed by a full rewrite, every time, and the gap compounds across a working week.

The five-part pattern

Nearly every effective workplace prompt contains five things, and most failed prompts are missing three of them. Role: who the AI should act as — not magic, just a compression of expectations (‘you are an experienced B2B copywriter’ carries a century of context in nine words). Context: the situation the response lives in — who the audience is, what happened before, what relationship is at stake. Task: the specific action, stated as a verb — draft, summarise, rewrite, compare, list. Format: what the output should look like — an email of no more than 150 words, three bullet points, a table with two columns. Constraints: the guardrails — what to avoid, what to include, what tone to strike, what not to invent.

Here is the pattern applied to a real, ordinary task. Vague version: ‘write a reply to this customer complaint’. Five-part version: ‘You are the customer-lead at a 40-person logistics firm. Context: a client’s Friday shipment arrived a day late; they’re annoyed and have CC’d their own boss. Task: draft a reply to their email below. Format: under 120 words, no corporate hedging, direct apology first, then the concrete fix, then one sentence on prevention. Constraints: don’t promise compensation beyond what I’ve approved (a waived invoice for the failed leg); don’t blame the courier; keep our existing friendly-professional tone.’ The second prompt takes perhaps forty seconds longer to type and produces a draft that is, in the ordinary case, genuinely sendable after a light edit.

How much context is enough

More context is not automatically better. The model can only attend to so much before instructions blur, and every sentence you add dilutes the ones that matter. The working rule: give the model everything that would change the response, and nothing that wouldn’t. For a customer reply, that means the relationship, the incident, the tone you want and the limits of what you’ll offer — not the client’s whole account history. For a document summary, it means the document plus the question you actually need answered (‘what did we decide about pricing?’), because a summary aimed at no question is a summary aimed at nothing.

There’s a confidentiality edge to this principle, and it cuts the same way. The less context you paste, the less you expose — so the discipline of giving only what changes the response doubles as the discipline of sharing only what’s safe to share. Often you don’t need to paste the sensitive data at all: describe it (‘a three-year contract worth about €200k with a termination clause in section 8’) and the model can do the work without ever seeing the real figures. We return to this in the data-safety guide; here it’s enough to note that good prompting and safe prompting are the same skill wearing different clothes.

Format: the most underused lever

If you take one improvement away from this guide, take this one: specify the output format, always. Unspecified format arrives as a wall of prose, because prose is the model’s default register. Specified format arrives shaped for its job. ‘Give me the answer as three bullet points I can paste into a slide’ produces something different from ‘write a 200-word paragraph for the report’, even when the underlying content is identical. Format instructions are also where length lives — ‘under 150 words’ does more to prevent bloat than any amount of asking for ‘concise’, which the model apparently defines as ‘anything under a page’.

The professional trick is to reuse format specifications. If every Monday you need the meeting notes as ‘decisions / actions / open questions, each with an owner and a date’, write that specification once and keep it — in a prompt library, a saved snippet, a team wiki. The specification becomes house style: every summary in the company arrives in the same shape, reviewable at a glance, comparable across weeks. This is the point where prompting stops being a personal trick and becomes an organisational capability, and it costs nothing but the discipline of writing the format down.

Iteration: the two-move steer

Beginners treat a prompt as a lottery ticket: type it, read the output, and either accept the prize or throw the ticket away and buy another. Professionals treat a prompt as a steering wheel. The first output is not the answer; it’s the model showing you its current understanding. Your job is to correct that understanding in one or two short moves — ‘shorter, and cut the exclamation marks’; ‘good structure, but lead with the apology, not the explanation’; ‘now the same thing for a reader who wasn’t in the meeting’ — rather than starting from scratch with a newly-worded plea.

Two moves is the working budget. If the draft isn’t converging by the third instruction, the problem is almost always missing context in the original prompt, and the efficient fix is to stop steering and supply the missing piece (‘forgot to say: the client has already threatened to leave; urgency is real’). Knowing when to stop steering and re-brief is itself part of the skill — it’s the difference between a five-minute exchange and a twenty-minute argument with a machine that is, throughout, perfectly willing to help.

The four workplace patterns

Underneath the infinite variety of tasks, most professional prompting falls into four reusable patterns. Rewrite: ‘here are my rough notes; turn them into X’ — the single highest-value pattern in office life, because everyone thinks in rough notes and ships in polished prose. Extract: ‘here is a document; pull out the decisions, dates and numbers’ — the pattern that turns reading a 40-page contract into reading ten bullet points. Compare: ‘here are two (or five) things; put them in a table by these criteria’ — the pattern that makes decisions reviewable. Role: ‘answer as a [tax adviser / skeptical customer / non-technical reader] would’ — the pattern that borrows a professional perspective to test your own work.

Learn the four patterns and most of your prompting life becomes fill-in-the-blank. Each pattern has a fixed instruction core (‘turn these notes into…’, ‘extract every commitment with its owner…’) and a variable slot where your specific task goes. Teams that standardise on the patterns get a second benefit beyond speed: consistency. When everyone extracts with the same instruction, everyone’s extractions look alike, and a colleague can read yours without a decoder ring. That consistency, not individual cleverness, is what ‘our people know how to use AI’ should mean.

The failure modes, so you can skip them

Five mistakes account for most disappointing prompts. The vague verb — ‘help me with this email’ gives the model no action to take; make the verb concrete (draft, rewrite, shorten, soften). The buried audience — if the reader matters, say who they are in the prompt, not in your head. The length vacuum — unspecified length defaults to too long; specify. The kitchen sink — twelve instructions in one prompt means two get followed and ten get sampled; split into a sequence when the list grows. The acceptance pass — reading the output with your eyes instead of your judgement, so the confident error ships because the sentence was fluent. The last one isn’t a prompting failure but a review failure, and it gets its own guide — but every prompting habit in this one is designed with it in mind: the better your brief, the smaller the gap between what the model wrote and what is true.

A note on tools, and on teaching this

Everything in this guide is tool-agnostic, deliberately. The five-part pattern works identically in ChatGPT, in Microsoft Copilot inside your Office apps, and in whatever assistant your tools ship with next quarter — because the pattern is a description of a good brief, not a sequence of interface clicks. That durability is the quiet argument for teaching prompting as a skill: tools change monthly, but ‘give the model a proper brief, specify the format, steer in two moves, and verify the result’ has been stable through every major model release so far.

It’s also, not incidentally, the module we teach first to every team. Watching thirty seconds of a three-minute video demonstrate the rewrite pattern on a real email does more than any amount of reading — then the task makes each person run the pattern on their own actual work, and the quiz checks the judgement. If your company would rather build this skill properly than by folklore, that’s Prompt Writing Essentials, included in every ClutterLearn subscription. But the guide above is the whole method; the module is just the fastest way to make it stick.