The way almost everyone uses AI today, and why it makes you the bottleneck

Let me describe a cycle I'm sure you'll recognize. You open ChatGPT or Claude, send a prompt, wait, get a response. Then you read it, evaluate it, realize it didn't come out quite the way you wanted, think about what needs fixing, and send another prompt. And repeat. And repeat. Until it's the way you expect.

Notice what's happening here? You're a step in the process. The AI only works when you press a button. It's as if you'd hired a brilliant employee who only acts when you ask, and just stands there waiting for your approval to continue.

In other words: you've become the AI's babysitter. And that creates two problems.

Problem 1: you're the bottleneck. The AI works for about 30 seconds and you spend the rest of the time reading, checking, and writing the next instruction. If you step away from the computer, the work stops. Nobody works without you there.

Problem 2, the worse one: you're also the one checking the work. And if you use AI, you know: it loves saying it did what you asked. It'll say the text turned out great, the site is perfect, the errors are fixed, and when you go check, it did half of it. So on top of babysitter, you've also become the quality inspector.

Meanwhile, the best AI users in the world are doing something completely different: they set up the loop once, put the AI to work, and go do something else. That's the difference between people who send prompts and people who build loops.

So what actually is a loop?

The Claude Code team put together a simple definition.

"A loop is the AI repeating work cycles until it hits a specific success criteria."

Claude Code team

Let me break that into two parts, because this is where everything lives.

Part 1: "repeating work cycles." Remember how in the prompt model, you were the one evaluating the result? In a loop, the AI does that itself. It executes, checks whether the work matches the criteria you defined, and if something's missing, it uses that feedback to improve itself, repeating the cycle until it judges the quality acceptable. All on its own, without you pressing a single button.

Part 2: "until it hits a success criteria." The success criteria is the answer to the question: how do we know the AI is done?

Here's a silly but perfect analogy: a regular washing machine. When does it stop? Not when your clothes are clean. It's when the one-hour cycle ends. It stops based on effort, not outcome. Whether the stain came out or not is your problem: you're the one who opens it, looks, and runs it again.

Now imagine a smart washing machine. You tell it: "wash until every stain is gone." It washes, checks the clothes, still sees a stain, washes again, and again, until it judges them clean the way you asked. That's a loop: you point the AI at the outcome, not the effort.

And here's the real shift: the work of loop engineering is far more about defining the success criteria than about executing. You stop being part of the flow and become the person who designs the flow. You stop giving orders and start defining verifiable goals, things like "only stop when every test is passing," "only stop when the page loads with zero errors," or "only stop when the text is under 500 words and covers these 5 points."

The 3 types of loop you can build

1. Turn Loop

This one you're probably already using without knowing it. Every time you send Claude a prompt, under the hood it runs a mini-cycle: gathers context, executes, checks its own work, and hands it back. That's the turn loop.

The problem is that self-check it does on its own is pretty weak. And you can improve it a lot by teaching the AI to review its own work, your way. The perfect tool for this is skills, that "instruction manual" that teaches the AI to work the way you work.

Think about the mental checklist you run to judge whether a result is good. You document that process inside a skill, and the AI itself starts checking against your method. A real example of mine: before handing me a YouTube script, the skill checks that it's under 3,000 words, that it cites the sources I used, and that there isn't a single word that makes the text sound AI-written. It only hands it to me once it clears those criteria.

And no, you don't need to know how to code for this. Just ask: "help me create a skill for..." and the AI walks you through the whole process.

2. Goal Loop

Claude Code has a command called /goal. Instead of giving an instruction, you give a goal plus a success criteria, and it stays in a cycle trying, checking, and fixing until it hits the goal, or until it reaches a retry limit you set.

The big difference from the turn loop is who verifies. Here it's neither you nor the AI itself: a second evaluator model comes in. That evaluator takes your criteria, looks at what the first AI did, and points out what still needs work. The loop only ends when the evaluator approves. It's one AI holding another accountable for doing it right.

3. Time Loop

This one is perfect for recurring tasks, the ones you do every day or every week. In Claude you use two commands: /loop and /schedule. The most important difference: /loop runs on your computer, so it needs to be on, and /schedule runs in the cloud, so it keeps working even with your laptop closed.

A few practical examples: every morning at 8am, read Slack and prepare a summary of what you missed; every 30 minutes, check whether a live project is still working and flag it if something broke; every Monday, mine for video ideas and bring back a list of topics. That's exactly the kind of loop I used to put together this week's video.

One important tip: don't sleep on tokens. When you say "work until you can't anymore," the AI will do exactly that, and it can burn through your tokens doing it. Especially with /goal, always set a retry limit, something like "try at most 5 times." That puts a ceiling on it and keeps the loop from running forever.

Recap

Here's the core message: most people use AI to be a little more productive, but stay the bottleneck of the operation. People who build loops step out of the operation, and that's where productivity unlocks at a level no other way of using AI reaches.

My recommendation for getting started: pick one recurring task you already do and start simple. Build a skill that defines a quality criteria, or schedule a loop with /loop or /schedule. Start with just one. Once you try it, you won't want to go back.