Picture a sales manager tapping a prompt into a phone between meetings: “write a quote email.” Three seconds later they have a cheerful, generic paragraph about “quality service” that could be selling software, storage or stationery. They drop in the client’s name and send it, then wonder why it reads like nobody in particular wrote it. That is not the AI being bad at its job. It is one of a small handful of habits that catch out almost everyone, and each has a straightforward fix.

If you have not read our guide on how to write a prompt that actually works yet, start there. It covers the eight things that make a prompt land properly. This is the follow-up: the ways people trip up in practice, with a before and after for each.

Being too vague

“Write a quote email” is a hint, not an instruction, so the AI fills the gaps with the blandest version it can produce and gets blamed for the result. One extra sentence fixes it: “You are an account manager in Stockport quoting for a support contract, £600 to £900 a month depending on cover, two-week onboarding, sign off warm and direct, like writing to somebody you know, not writing for a call centre.” Same tool, same three seconds, a different result, because now it has something real to work from.

Assuming it remembers last week’s chat

A cafe owner spends a Tuesday getting the AI dialled in on her specials board copy. The following Monday she opens a fresh chat and gets something flat back, and assumes the tool has gone downhill. It has not. Every new chat starts from nothing, whatever happened last time. Keep your basics, your voice, your prices, your usual phrases, saved somewhere you can paste in at the start of each chat. Our Prompt Builder tool is built to make that quick.

Accepting the first draft as finished

An account manager asks for a renewal quote, gets a smooth reply, and pastes it straight to the client. It later turns out the AI got the VAT wrong and quoted a support rate two years out of date, because it was guessing at numbers it was never given. Fluent and confident is not the same as correct, especially with prices, dates and regulations. Read every number as if a brand new starter wrote it, because in a sense one did, and check it against your real price list first.

Skipping tone and format, then being disappointed it “doesn’t sound like us”

Left with no steer on tone, AI defaults to a polished, faintly corporate voice. A marketing lead asks for a LinkedIn post about a finished project and gets “we are delighted to showcase” when the company’s usual posts are two lines and a straight fact, because nobody told it what the company sounds like. Paste in a real post or email you wrote last month and say “match this tone and length.” It is one of the most reliable fixes there is.

Asking for one giant output instead of steps

A marketing lead wants a new website written and asks for the homepage, services page, about page and FAQs all in one prompt. What comes back is shallow everywhere, because the AI is doing four jobs’ worth of thinking in one pass with no chance to check any of it. Break it into stages instead, the way you would brief a member of staff: headline first, check it, then services, then about. Each step gets proper attention, and problems get caught early rather than buried on page four.

Forgetting you are a UK trade talking to UK customers

An operations lead asks for “a standard quote template” without saying where the business is based, and gets back dollar signs, sales tax and contract terms from a different country’s law entirely. It reads plausible if you do not look closely, which is the danger. Say so up front: UK business, prices in pounds, VAT at the current rate, and the actual rules that apply to you, whether that is UK GDPR for customer data or HMRC guidance for how you treat a charge.

Sending it straight out the door with no human check

This is the big one. A marketing lead puts AI-drafted terms and conditions on the website with no review. An operations manager copies AI-written compliance wording straight into a client document. Both assume that because it reads well, it is safe to send. Anything touching money, dates, regulations or legally binding wording needs a human, ideally you, reading every word before it reaches a customer, a solicitor or the taxman. Treat every AI-facing document like a junior colleague’s first draft: often good, never sent unread.

Where to take this next

Go back to the pillar guide, how to write a prompt that actually works, if you have not worked through it yet. For a structured walk through these habits, our course Writing prompts that work takes you from a blank chat to prompts you would trust with a real customer, and our Prompt Builder tool does much of the heavy lifting for you, keeping your business context ready so you never start from zero again.

If you would rather work through it with people in a room, that is what The £1,000,000 AI Blueprint in Leigh is for. Have a look at our upcoming events. See the event page for current booking details.