My AI Rewrote My Script Nine Times. I Only Asked Once.
Loop engineering is the free AI skill nobody told you about.
I’m watching my screen and my AI is on its fourth pass of my latest YouTube script.
It’s not frozen. It’s not broken. It’s rewriting… the whole thing… because it scored itself a 6.8 out of 10 on “hook strength” and that’s below the threshold I set.
I didn’t tell it to rewrite. I told it what done looks like.
It rewrote itself nine times. Scored every pass against nine different criteria. Calculated the overall average each time. And it stopped… cleanly, on its own… when it hit 9 out of 10.
I set this up once. Took me about fifteen minutes. And it’s completely free. The loop is built into the model already. You don’t need a special tool, a plug-in, or a paid upgrade. You just need to know what to ask for.
What’s actually going on here
For about two years, the way you got something useful out of AI was prompt engineering. You learned to write careful instructions. You figured out which words made ChatGPT or Claude give you better answers. “Act as a...” “Step by step...” “You are an expert in...”
That worked. But it assumed something specific - that you’d be sitting there the whole time, typing one thing, reading what came back, typing the next thing. The AI is a tool and you’re holding it, one turn after the other.
That part is kind of over.
Boris Cherny - he’s the head of Claude Code at Anthropic - put it bluntly: “I don’t prompt Claude anymore. I have loops running that prompt Claude. My job is to write loops.”
Hold that thought.
The one-line version
Loop engineering is this: instead of prompting your AI and checking the result yourself, you tell the AI what “done” looks like and let it check its own work… over and over… until it meets your standard.
That’s it. That’s the whole discipline in one sentence.
The shift sounds small. It isn’t.
Prompt engineering asked: How do I talk to this machine? Loop engineering asks: Do I actually know what I want?
And honestly... most people don’t. Not precisely enough. Which is exactly why the loop either runs forever or quits too early. The engineering isn’t the AI part. It’s the knowing-what-good-looks-like part.
What this looks like in practice
Here’s exactly how mine works. I write YouTube scripts, so I built a loop around that… but this pattern works for anything. Emails, marketing copy, blog posts, social media, proposals. Anything you’d normally write, read back, frown at, and rewrite yourself.
I gave my AI nine evaluation criteria:
Hook - Does the first 30 seconds land on something specific, not a generic intro?
Voice - Does it actually sound like me, not a content creator performing?
Specificity - Real numbers, real names, real tools. No “many experts say.”
Story structure - One clear through-line. Every detour comes back.
Honest moment - At least one beat where I admit something didn’t work or took me three tries.
Actionable value - Does the viewer walk away knowing something they can actually use?
Verdict clarity - Does it take a clear position, or hedge into “it depends on your use case”?
Pacing - Pattern interrupts, section breaks, moments that pull you through the middle.
Length discipline - No padding, no filler. The script is as long as the idea needs and not a word longer.
Each one scores 1 to 10. The overall average has to hit 9 out of 10 before the loop stops.
The first pass usually comes back around a 6 or 7. Weak hook. Too vague in the middle. The pacing drags somewhere around minute eight. By pass four or five, it’s tightened up noticeably. By pass eight or nine, it’s genuinely good… and it got there without me typing a single correction.
Here’s the thing nobody’s talking about. The AI didn’t get smarter between pass one and pass nine. Same model. Same capabilities. Same context window. What changed is it had a clear target to measure itself against. The loop didn’t make the AI better. My criteria did.
You can do this right now
Here’s the prompt template I use. Copy it. Paste it into ChatGPT, Claude, Gemini - any of them. Fill in the blanks for whatever you’re working on.
I need you to [YOUR TASK].
Before you give me the final version, evaluate your work against these criteria. Score each one from 1 to 10:
1. [WHAT MATTERS MOST] 2. [YOUR SECOND CRITERION] 3. [YOUR THIRD CRITERION] (add as many as you need)
After scoring, calculate the overall average.
If any single criterion scores below 7, or the overall average is below [YOUR THRESHOLD]: - Identify which criterion scored lowest - Explain in one sentence why it scored low - Rewrite the entire output to address it - Score again
Keep going until every criterion hits at least 7 and the overall average reaches [YOUR THRESHOLD]. Show me the scores for each pass so I can see the improvement.
That’s it. Fifteen minutes of thinking about what “good” means for your specific task, and the AI handles the iteration.
The harder question… and the more useful one… is choosing the right criteria. Three to five is plenty for most tasks. Nine is for something you do every week and care about deeply. The criteria are the product. The loop is just the delivery mechanism.
This is already everywhere
Greg Isenberg released a video yesterday where he and Elie Steinbock walked through using loops to run entire businesses. An SEO loop that connects to Google Search Console, runs once a month, pushes your rankings up automatically… for a few dollars per run. An ad loop that tests creative variants. A product feedback loop that reads customer reviews, checks your analytics, and prioritises what to build next.
The pattern is always the same. Give the AI a task. Give it an objective metric. Give it a stop condition. Let it improve on a schedule.
Eric Ries wrote “Build, Measure, Learn” in The Lean Startup fifteen years ago. Toyota had the same cycle decades before that. Loop engineering is that idea… but the AI does the building, the measuring, and the learning. Your job is to define what success looks like.
Once you see it, you can’t unsee it. Every repeatable task in your work… every email you write the same way, every report you format on Fridays, every social post you agonise over for twenty minutes… is a loop waiting to happen. You just have to be honest enough about your own standards to write them down.
The skill that replaced prompt engineering isn’t technical. It’s the oldest skill there is… knowing what you actually want… applied somewhere it finally has leverage.
Cheers, Jagger