Stop hunting for prompt formulas. Start closing the loop.

There's a whole industry telling you that getting good results from AI is about prompt formulas. Frameworks with acronyms. Magic phrases. Courses that teach you to say "act as a world-class marketing expert" before every question.

Most of it misses the point. And the problem becomes obvious when you look at the shape of good AI work.

It isn't a straight line from prompt to output. It's a circle. It starts with you, passes through the machine, and comes back to you. And it keeps going round until the work is genuinely yours. You don't need a prompt formula. You need a working loop.

Here's how it goes round.

It starts with context

The quality of what comes out is set by the quality of what goes in. Not clever wording. Context. Who is this for? What does good look like? What's the history? What are the constraints?

You already know how to do this. You do it every time you hand work to someone else. Brief AI like you'd brief a capable new starter, and you'll get a capable first draft. Brief it like you'd shout across the office, and you'll get something plausible and wrong.

The fastest shortcut is an example. Don't describe the report you want. Hand over last month's report and say "like this, with the new figures". An example answers fifty questions at once. Describing it answers three. If you want the mechanics of that, we set them out in how to write a prompt that actually works.

And here's the trade-off nobody tells you: you pay for output quality up front or in iterations. Never neither. A good brief gets you there in two exchanges. A lazy one takes six. Same destination, different cost, and across a team doing this daily that difference is hours a week.

Then demand the truth

AI models lean towards agreeing with you. They're trained on human feedback, and humans reward politeness. Show one your proposal and ask what it thinks, and it will tell you it's strong. They also follow your framing: present something as good work and ask for feedback, and they'll reinforce your assumptions rather than test them.

For anything that matters, whether that is a sales proposal, a campaign or a client email, you have to ask for the hard version. And not "be brutally honest", which just changes the tone. Ask a specific question with teeth:

"What would make a procurement panel score this down?" "Which claims would a sceptical finance director ask us to prove?" "Where would a rival undercut this?"

And take your thumb off the scale. "Review my draft" tilts the answer before it starts, the same way telling a reviewer which option you prefer does. Present the work neutrally and let the critique land where it lands.

For the big stuff, get a second opinion

You wouldn't send your most important proposal out with one reviewer. There is work where that matters: the tender, the campaign, the difficult letter. For those, run the same document and the same hard question past a second model. Claude, ChatGPT and Gemini have different strengths and different blind spots, and a second pass takes ten minutes.

Where they agree something's weak, that's a useful signal. Not proof, because they share some of the same habits, but worth acting on. Where they disagree, pay closer attention: that's not the system failing. It's the system finding a genuine judgement call, and a judgement call needs a person who's accountable for it.

Every stage so far has been about getting better advice. This last one is about who carries the can.

The circle closes with you

Here's the thing the prompt merchants never mention. AI can only work with what it was told. It doesn't know your client flagged budget pressure on Tuesday's call. It doesn't know this is a renewal, not a new bid. It may not recognise that essential context is missing. It can give you a fluent, confident answer without knowing it has answered the wrong question.

You're the only one in the loop who can spot what the machine was never given, and who else needs to check the parts you don't hold either.

But closing the loop doesn't just mean accepting or rejecting the answer. You tell it what it missed. You correct what it misunderstood. You challenge an assumption, add a constraint, ask it to go again. Your judgement becomes new context, and the work goes round the circle once more.

You keep going until the result isn't merely something AI produced. It's something you understand, believe, and are willing to put your name to. Because when it goes out under your name, "AI wrote it" isn't a defence. Not to your client, your board or your regulator. It never was for work someone else drafted, and it isn't now.

AI advises. You decide. You own it.

The loop

Look at the shape of it. Your judgement goes in as context. The machine does the heavy lifting. Your judgement comes back as criticism, correction and direction. Each pass closes the gap between a plausible answer and work you'd actually stand behind.

Brief. Produce. Challenge. Refine. Decide. Round again until you understand it and are willing to own it.

Three of the five are yours. That is the point: the machine is one step of a process you are running, not the process.

The people getting the best results aren't the ones with the fanciest prompts. They're the ones running the tightest loop, and every part of it is a habit you already have, pointed at a new colleague.

That's what we teach: how to turn AI from a one-shot answer machine into a dependable working process. One day, hands-on, using your own documents, not invented exercises. You'll leave with a way of working you can use the next morning.

Written with Claude, then steered, cut and checked by me. I answer for it.

Finlay Carmichael

Finlay Carmichael is the founder of The AI Lab, helping teams put AI to practical use through hands-on training and clear advice. He brings nearly 30 years' experience building and leading a Microsoft Business Solutions partner, working with hundreds of clients across the UK. Request a call