This is how most of us work with AI. We write a prompt. We read what comes back. It is 70% right, so we fix it, or  explain what was wrong and ask again. It comes back better. We fix it again. Five or six rounds later, there is something you can use.

It works. We shipped a lot of content this way. But look at what it actually is: a person and a machine taking turns, and the person is the slow part. A lot of imperfect prompts and lost tokens.

There is a better way, and once you see it you cannot unsee it.

The old way puts you inside every turn of the cycle. Prompt, result, fix, result, fix. You are the engine.

The new way puts you in front of it, once. You define what a good result looks like and how to check it. Then the AI does the work, checks its own output against your standard, sees what is missing, and goes again. It keeps looping on its own until it hits the bar you set. You come back to a finished result, not a first draft.

Note! Work on your prompts! Bad prompts and loops means a loooot of tokens that go to waste.

The difference is where the human sits. In the old way you are inside the loop, doing a turn every time. In the new way you build the loop, and then you step out of it.

People who write code have been calling this loop engineering. It is not only for engineers. It is just as useful for a newsletter.

I did not prompt my way to this issue.

I gave one instruction with the checking built in:

It drafted, graded itself, found what was weak, and rewrote, three times before I read a single word. My 50 minutes went to the last version instead of the first one. Same standard, applied earlier, by the thing doing the work.

The rule was in the instructions the whole time. But on a single pass, while the AI is busy writing, checking gets skipped. The loop pulls the check out into its own step, where it actually runs. It is the difference between "keep my voice rules in mind" and "now go check this draft against them." 

That is the whole idea. You are not writing the content. You are writing the definition of good and the test for it, once, and letting the work run against that test until it passes.

This issue is brought to you by HubSpot.

The obvious question is how far this goes. What does it look like when the loop is not a newsletter but an entire marketing organization, with a sales team, a pipeline, and money on every decision?

That is the book I want to read.

Loop: Outlearn. Outmarket. Outgrow. comes from HubSpot CMO Kipp Bodnar and SVP of Agentic GTM Kieran Flanagan. It is what they learned rebuilding HubSpot's own marketing organization for the AI era, and the argument is that inbound was the operating system for the search era, and this era needs a different one. Coming from the company that invented inbound, that is worth my time.

The book ships September 22. Pre-order now and it is $26, the price of the book, with about $335 of things around it: a ready-to-use prompt library, the first three chapters early, a private three-episode podcast series, a Claude Skill built from the book, and a live Q&A with Kipp and Kieran.

#HubSpotMediaPartner #Ad

Pick one thing you make on repeat. A newsletter, a set of captions, a first-pass script.

Write down two things: what a good one looks like, and how you would check it. Then hand both to the AI and ask it to keep going until its own work passes the check. 

There is an even more advanced version, where the loop improves itself every run. It is what Boris Cherny, the creator of Claude Code, describes: agents that keep refining their own work in the background, getting better each time. That one needs its own email. Because that’s a self-improving loop and AI’s recursive improvement is the hottest topic right now.

Improvement happens when you give feedback to AI and it learns from your feedback, basically learning about your taste. Now think about where your taste lives. Is it zoom calls? Start recording them with Granola and connect Granola to your Claude. Do you give feedback on Slack? Add a bot to your chat, a bot that will remember and process what you like and don’t. And then ask your AI to constantly learn and update itself on your taste. Launch that loop. And we’ll dive deeper into recursive improvement in my next emails. 

In order to start spending less time on manually prompting AI, you need to start spending less time on the output, and more on quantifying and describing what you consider great and setting benchmarks. And this is how you become the manager of AI that everyone on social media is talking about. 

Marina 💜