The prompt patterns I actually use for business work

Most consultants would put this guide behind a discovery call. Here is the method in full, because the honest truth is that knowing it is not the moat. Doing it is.

That is the whole bet behind this series. I give away the keys of what I do because people will not do it. They will read this, nod, agree with every word, and tomorrow they will type "write me a marketing email" into a chatbot and complain the output sounds like everyone else's. More than half of UK SMEs are now using AI, according to the British Chambers of Commerce, and most of them are getting average results because they are writing average briefs. If you are one of the few who will take this and actually run it for a month, it will change how you work. If not, the bottom of this page tells you what to do instead.

One thing before we start. We have a full guide to writing business prompts that covers the theory: context, task, format, constraints, examples. That one is written for someone starting cold. This one is the practice. These are the six patterns I use daily, in roughly the order I built them.

1. Write the context once, then never again

I do not retype who the business is, who the reader is and how we sound every time I open a chat. I wrote it once, as a context block, and I paste it at the top of anything that produces customer-facing words. If your tool supports custom instructions or projects, it lives there and you stop pasting too.

Here is the shape, for a made-up business:

You are writing for Harborne Joinery, a six-person fitted furniture workshop in Birmingham. Customers are homeowners spending £4,000 to £20,000 on alcove units, wardrobes and home offices. They care about craftsmanship and being treated like adults, not "design journeys". Tone: plain, warm, confident. We never discount and we never hard sell. British English throughout. Sign off as Pete.

That is under 80 words and it changes everything downstream. Mine took about twenty minutes to write and I have reused it hundreds of times. The return on that twenty minutes is the best in this whole guide.

2. Show it one real example of done

Adjectives are the weakest instruction you can give a model. "Professional but friendly" means a different thing to every business on earth, so the model averages across all of them and you get beige. One real example pins it instantly.

The pattern is two lines:

Here is a quote follow-up that won us the job last month: [paste it]. Match its tone and structure, not its content.

No example to hand? Make one. Take the model's first attempt, edit it until it sounds like you, then save the edited version as the reference for next time. You do this once per task type and never again. This is exactly how we built Bidwell: it drafts tender answers from a business's real past bids, not from adjectives, because a model shown a genuine winning answer writes in your voice and a model given adjectives writes boilerplate.

3. Constrain the shape until it cannot ramble

Left alone, every model produces output that is too long, too bulleted and too pleased with itself. So I specify the shape harder than most people specify the task:

Output exactly this: a subject line under seven words, then a body under 110 words in three short paragraphs, then a one-line ask ending in a question. No bullet points. No "I hope this finds you well". Nothing after the sign-off.

The load-bearing words are "exactly this" plus the counts. Models obey numbers far better than they obey vibes. "Keep it short" gets you 300 words. "Under 110 words" gets you under 110 words.

4. Make it cite or admit it cannot

For anything factual, research summaries, competitor notes, stats for a post, I add this:

For every factual claim, give the source URL. If you cannot find a real source, write "no source found" next to the claim. A gap is fine. A confident wrong number is not.

Then I click the links, because some will still be wrong. The point of this pattern is not that it makes the model honest. It is that it makes the lies visible. Most fabricated facts ship because nobody forced the model to show its working, and nobody checked. On this site, no number ships without a source, and that rule started as this prompt line.

5. Ban the words you hate

Negative constraints work, and almost nobody uses them. I keep a standing ban list inside my context block:

Banned words and phrases: leverage, seamless, journey, empower, unlock, elevate, delve, game-changer, "in today's fast-paced world". No em-dashes. No exclamation marks. If a sentence wants one of these, rewrite the sentence.

Your list will differ. The test for what goes on it is simple: every time you catch yourself deleting the same word from AI output, it goes on the list. Within a fortnight the output stops sounding like AI output, because the tells are exactly these words.

6. Iterate on the brief, not the output

This is the pattern that separates people who get compounding value from people who get one decent email. When an output is wrong, most people patch the output: "make it shorter", "warmer", "less salesy", ten rounds deep into a chat that is now juggling contradictory notes. I patch the output once, and then I move the fix upstream into the saved brief.

Concretely: one document per recurring task, the full prompt at the top, the best output below it as the example. Every time I make the same correction twice, the correction becomes a line in the prompt. The next first draft starts where the last edit ended. After a month, every recurring task in the business has a brief that produces a sendable draft on the first attempt.

The brief is the asset. The outputs are disposable. Get that the right way round and you have a prompt library that is worth more than any course, and, when a task earns it, that library is the exact specification for an automated system. Every system we build for clients starts life as somebody's well-worn prompt.

Where this goes wrong

I have watched each of these happen, including to me.

• Confidential data in the context block. Client names, financials, anything personal: anonymise it before it goes into a general-purpose chatbot, and check the tool's data terms if you handle customer data at scale. The ICO's AI guidance is the UK reference.
• Prompt hoarding. Fifty saved prompts you never reuse is a stationery cupboard, not a system. Five briefs you run weekly beats it every time.
• The context block rots. Your business changes, your pricing changes, your block does not, and suddenly every output is confidently out of date. Reread it monthly. Takes two minutes.
• Trusting pattern 4 without clicking. "Cite your sources" reduces fabrication, it does not eliminate it. The checking is the job.
• Iterating past the point of return. If you are six rounds deep and the output is still drifting, the chat has accumulated contradictory instructions. Start fresh, paste in the refined brief plus the best draft, carry on. You lose nothing.

What it actually takes

Here is the honest bill. Twenty minutes to write the context block. A few minutes per recurring task to write the first brief. Then the part that filters out almost everyone: the discipline to keep editing the briefs instead of the outputs, every working day, for weeks, when it would be quicker today to just fix the email by hand. There is no skill barrier in any of this. There is a stamina barrier, and it is why I can publish the entire method without losing a single client.

If you would rather skip the stamina part and have someone build the briefs, the library and the systems they turn into, book a consultation and we will just do it.