Practical AI · Episode 46≈ 6 min

How to stop prompting AI and start writing loops

The framework is Dharmesh Shah's. The part where I tried it in my own business, that's the story.

A first-person walkthrough for the show. Framework credit: Dharmesh Shah, co-founder of HubSpot, from his simple.ai newsletter "How to Write Your First AI Loop."

The hook

Here's something that changed how I work this week, and I think it can change yours. Dharmesh Shah, the co-founder of HubSpot, writes a newsletter called simple.ai. This week's issue was titled "How to Write Your First AI Loop." I read it twice.

The line that stopped me came from Boris Cherny, the engineer who created Claude Code:

"I don't prompt Claude anymore. I write loops, and the loops do the work. My job is to write loops."Boris Cherny, quoted in Dharmesh Shah's newsletter
Dharmesh's framework (all credit to him)

What a loop actually is

Most of us use AI like a vending machine. Type a request, get an answer, type the next one. That's prompting. A loop is different, and the whole framework here is Dharmesh's. A loop has three parts.

ObjectiveWhat "done well" looks like, in one sentence.
MetricHow the AI scores its own work, so you're not reading every word.
BoundaryHow far it runs on its own before it stops and checks with you.

His sharpest point: there's a loop that runs and a loop that learns. A loop that runs does the same thing every day. A loop that learns has a feedback signal wired in, so it gets better every time. Aim for that one.

My idea

Could a non-developer actually build these?

Here's where I come in. I'm not a developer. I think in outcomes, not code. So when I read this, my first thought wasn't "nice theory." It was: could I actually build loops for my own businesses? My recruiting firm, my sales-training startup, this show. Real work, today, not someday.

The prompt I ran

One simple instruction

So I sat down with Claude and gave it one instruction. No code, just plain English.

What I typed"Read this Dharmesh email. Then give me five loops I could run for each of my businesses, using the way I actually work."

That was the whole thing. Read this. Apply it to me. Five per business.

The result

A playbook came back

Twenty loops. Five for each of my four businesses. A loop to find and score candidates. A loop to keep my content on-message. A loop to turn every episode into something AI engines will actually quote. Each one had that same shape: a goal, a way to grade itself, and a line where I still decide.

Example 1
The one I tried first

What happened when I ran it for real

Then I did the part most people skip. I picked one and actually ran it. I chose a loop to reach back out to deals that had gone quiet. I won't get into the names. Here's the part that got me.

The loop drafted the messages. But I'd given it a boundary: check my real email history before anything is final. And when it did that, it caught that its own first drafts were wrong. It had assumed a backstory that didn't match what actually happened with those people. So it rewrote them, accurately, before I ever saw a bad version.

That's the whole point. The loop didn't just do the work. It checked itself against reality and fixed its own mistake. The judgment, "verify before it's ready," that part was mine. The loop carried it out.

Example 2
The loop I like even better

Then I used a loop to name this whole show

Here's the one that really shows it off, because you can watch every single pass. I run a loop to write the title and hook for each piece of the show. Same three parts, and the feedback signal is baked right in.

ObjectiveA title that leads with the viewer's value, not "what I did."
MetricSelf-scored 1–10 on a five-point rule. Rewrite anything under 9.
BoundaryIt hands me ten winners. I gut-pick.

The five-point rule it grades against: leads with viewer value, one clear promise, curiosity without clickbait, under about 60 characters, and it kills my banned framing ("watch what I built / I tried / my…"). That last rule is the feedback signal. It's what makes this a loop that learns my taste, not just one that runs.

The juicy part is watching a weak title climb to a strong one. Three real ones from this very episode:

"Watch How I Built AI Loops For My Business" 2/10 "How To Build An AI Loop" 7/10 "Make AI Do The Work While You Sleep" 10/10
"AWS Adds AI Traffic Monetization To WAF" 2/10 (jargon, zero viewer value) "Get Paid When AI Reads Your Website" 10/10
"My Brain Is Fried From Using AI All Day" 3/10 (banned "my," all problem no fix) "Why AI All Day Fries Your Brain (And The Fix)" 10/10

Same shape as the first loop: it does the cranking, the judgment stays mine. It generated, scored itself against my rule, rewrote everything under 9, and handed me ten finished titles to choose from. See all ten title loops, scored and rewritten →

What I want you to try this week

Take one thing you already use AI for. Drafting emails, prepping for a meeting, summarizing a call. Write down three things.

  1. The objective. What does "done well" look like, in one sentence?
  2. The metric. How would it grade itself, without you reading every word? If you can't answer that yet, you just found the real work.
  3. The boundary. What can it do alone, and where does it stop and check with you?

That's a loop. The keystroke work goes away. The judgment work stays with you. That's the good news.

The close

Again, the framework is Dharmesh Shah's, from his simple.ai newsletter "How to Write Your First AI Loop." Go read it. Then come build one. I did, and I'm not a developer.