Prompt Engineering for Business: A UK Manager's Handbook

Under-specifying context. The difference between 'write a customer follow-up email' and 'write a 120-word follow-up email from a UK account manager to an SME client in financial services, confirming next Tuesday's renewal meeting, with a warm but professional tone and a single clear call to action' is enormous. Managers used to delegating work to skilled colleagues tend to under-brief AI because they assume shared context that doesn't exist. Train your team to think of AI as an intelligent but context-free colleague: role, objective, audience, constraints, format — every time. Short prompts are fine for throwaway tasks, never for anything that will be sent externally.

Yes, for any prompt used more than a handful of times. A prompt library — stored in Notion, Confluence, a GitHub repo or a shared folder — solves three problems: new hires reach competence faster, improvements compound across the team rather than getting lost in individual chats, and you get version control that lets you roll back prompts that degrade. Start with five to ten high-traffic prompts per department (customer response drafts, meeting summarisation, sales-email variants, research synthesis, report templates). Tag each prompt with who owns it, when it was last reviewed, and the model it was tuned for. Review quarterly.

Generate the first output with your context-dense prompt. In the same conversation, ask the AI to critique its own output against specific criteria ('list three weaknesses in that response from the perspective of a busy UK SME owner' or 'identify any claims that need evidence before this could be sent'). Use the critique to refine — either manually or by asking the AI to regenerate incorporating the critique. The critique step is where most of the value lies because it surfaces issues a tired reviewer would miss. Three passes is usually enough; more than five means the initial prompt wasn't strong enough and you should rewrite rather than keep iterating.

Structured output — JSON, a markdown table, a bulleted hierarchy — whenever the output is consumed by another system or reviewed under time pressure. Meeting summaries, competitor analyses, sales-call reviews, customer research synthesis and status reports all benefit. Prose is right when the output is itself the deliverable — customer emails, blog posts, marketing copy. The common mistake is asking for prose when a structured format would be faster to scan and easier to convert into action items. If your team spends hours reading AI-generated summaries, switch those prompts to structured output and measure the time saved.

Feed the model real examples. A prompt that says 'write in our brand voice' produces generic output; a prompt that includes three actual recent emails from your company and says 'match this tone, register and sentence length' produces output that's almost indistinguishable from the real thing. Build a short brand-voice block — 200–400 words of style guide plus three representative examples — and prepend it to any external-facing content prompt. Update the block quarterly. This single pattern is responsible for most of the difference between teams that say 'AI sounds generic' and teams that say 'AI sounds like us'.