Why Looping Is Better Than Prompting AI (And Why the Man Who Built Claude Agrees)

The Man Who Built Claude Stopped Prompting It. Here Is What He Does Instead.

Boris Cherny is the engineer at Anthropic who created Claude Code, which means he has spent more time thinking about how to get the best out of AI than almost anyone on the planet. Last year he said something that should have made every small business owner put down their tea and pay attention.

"I don't prompt Claude anymore. I have loops running that prompt Claude and figure out what to do. My job is to write loops."

The man who built the thing stopped using it the way most people use it. That is worth sitting with for a moment.

What You Are Actually Doing When You Prompt

Every time you open Claude or ChatGPT, type something in, and take whatever comes back, you are prompting. You get one answer. You read it, decide if it is good enough, and either use it or try again with slightly different wording and slightly more frustration.

It works. The same way sending one follow-up email to a customer who went quiet works. Technically you tried.

But here is the thing nobody tells you. When you prompt, you are treating the first attempt as the final version. And nobody who is serious about quality does that. Not chefs. Not solicitors drafting contracts. Not the mechanic at the garage who test drives the car twice before handing back the keys.

One shot at something that matters is not a method. It is a gamble.

What a Loop Does Instead

A loop does not give you one answer. It runs the same task multiple times, scores every version against a standard you set, and hands you only the best one. You never read the attempts that did not make it. You only ever see the winner.

Say you run a dental practice in Surrey. A patient attended a consultation for Invisalign three weeks ago and has not booked to proceed. You need a WhatsApp message that feels personal without feeling like a sales nudge, clear without feeling clinical, short enough to read between appointments.

If you prompt, you get one version. You read it, wonder if it is a bit much, send it anyway.

If you loop, the AI writes five versions of that message, scores each one on warmth, clarity, and brevity, and returns the highest scorer. You never saw the awkward ones. You sent the best one.

Same task. The difference is who is doing the quality control. When you prompt, it is you. When you loop, it is the AI.

This works exactly the same way whether you run a dental practice, a car dealership following up on a test drive, a physiotherapy clinic chasing a course of treatment, or a beauty salon trying to rebook a client who has not been in since Christmas.

Three Ways People Currently Use AI in Their Business

Beginners: take whatever comes back. No structure, no scoring, no standard. Results vary because the method varies. Most people start here and stay here longer than they should.

Intermediates: write better prompts. Give context, specify the tone, name the audience. Results improve. But you are still working with one attempt at a time.

Advanced: run loops. Set the standard before the AI starts. Tell it how many versions to produce, what to judge them on, and to show you only the winner. The AI does the quality control. You approve the result.

Most business owners in Ireland and the UK are somewhere between the first and second. Boris Cherny is firmly in the third. This post gets you there.

The Part That Decides Everything: Writing a Proper Goal

A loop is only as good as the brief you give it. A vague goal returns vague versions. No amount of running five attempts fixes a poorly written instruction.

A good goal has three things.

First, say exactly what you want. Not "write me a follow-up message." Instead: write a WhatsApp message for a customer who test drove a Ford Focus last Thursday and has not responded to our initial call. They seemed interested in the finance options. That is a brief. The first version is a guess.

Second, tell the AI how to judge its own work. Give it two or three criteria. For the car dealership: warmth, so it does not read like a debt collection notice. Clarity, so the customer knows exactly what to do next. Brevity, because nobody reads a long WhatsApp from a business they are not sure about yet.

Third, set the limit. Five versions. Scored on all three. Show me only the one with the highest combined score.

Three things. The entire quality of your output lives or dies on how clearly you write them.

The Template: Copy It, Change the Brackets, Use It Today

This works in Claude and in ChatGPT. Change what is in brackets to match your business.

You are a communication specialist for [type of business, e.g. aesthetic clinic / dental practice / car dealership / physiotherapy clinic].

GOAL: Write a [type of message] for [describe the situation specifically]. Tone: [warm and professional / reassuring / direct]. Length: [under 80 words / two to three sentences]. Channel: [WhatsApp / email / SMS].

SCORING: After each version, score it out of 10 on: Warmth: does it feel like it came from a person, not a system? Clarity: does the reader know exactly what to do next? Brevity: is every word earning its place?

LOOP: Write 5 versions. Score each one. Show me only the version with the highest combined score. Include the scores so I can see why you chose it.

For a skin clinic in Dublin: "aesthetic clinic," "follow-up WhatsApp for a client who attended a consultation for a course of peels but did not book," "warm and professional," "under 80 words," "WhatsApp."

For a dental practice in Manchester: "dental practice," "follow-up for a patient who attended an Invisalign consultation three weeks ago and has not responded," "warm and reassuring," "under 80 words," "WhatsApp."

Change the brackets. The structure does the rest.

One Extra Step That Makes It Twice as Good

A loop running on its own is useful. A loop running inside a Claude Project that already knows your business is something else.

Open Claude. Go to Projects. Create one for a specific task, something like "Customer Follow-Up Messages" or "Booking Reminders." In the Instructions field, write one short paragraph about your business: what you do, who your customers are, what tone you use. Clear beats long.

Then upload three examples of your best existing messages. Your best rebooking reminder. Your best lapsed customer text. Your best post-consultation follow-up. The AI reads them and uses your own standard as the benchmark.

Now run the loop inside the Project. Every version it produces will already sound like you. The loop then filters those and shows you the winner.

You are not writing from scratch. You are approving the best version of your own voice, produced in about forty seconds.

Why the Man Who Built Claude Does Not Prompt Anymore

Boris Cherny did not stop prompting because prompting is bad. He stopped because once you have seen what looping produces, prompting feels like ordering one dish from a restaurant and eating it whether you like it or not, when you could have tasted five and only paid for the best one.

Small businesses in Ireland and the UK are not short of AI tools. What most are short of is a method that actually turns those tools into consistent output rather than occasional good results.

Looping is that method. It works for a clinic in Cork trying to recover a missed booking. It works for a physio practice in Leeds chasing a referral. It works for a car dealership in Glasgow following up on a Saturday test drive.

The question is not whether AI can do this. It already can. The question is whether you are giving it one shot or five.

If you want to see how KARA our AI front desk can help your business, you can book a discovery call with us. No commitment. No pitch. Just the product working in real time.