Recursive Self-Improvement Loops for AI Prompting

it's called a recursive self-improvement loop. and once you understand it, you'll never prompt the normal way again.

I've been using this with amazing results lately. here's how it works:

most people prompt claude once and ship whatever comes back. a recursive skill works differently.

it generates, evaluates its own output against a scoring criteria, diagnoses what's weak, rewrites, and re-evaluates. it keeps looping until it actually hits the bar.

Example

shared an image ad skill that scores every concept against 10 criterias like thumb-stop power, curiosity gap, emotional trigger, persona recognition. if it scores below 9/10, it doesn't stop. it tells you exactly why it failed, rewrites the concept, and re-scores. over and over until it passes.

The Same Pattern Works for Everything

• email sequences: generates the full drip, then stress-tests every subject line, hook, and cta against the same loop.

• video hooks: evaluates the first 3 seconds against retention psychology. does it create a curiosity gap? does it pattern-interrupt? if a distracted scroller would keep scrolling, it rewrites and retests.

• positioning: stress-tests your angle against competitors, skeptical buyers, and market saturation. if a rival cmo could rip it apart, it finds the gap and repositions until the angle is unattackable.

• seo content: evaluates against search intent, depth, and competitive gap. if a top-ranking piece already says the same thing, it finds the angle they missed and rebuilds around that.

the pattern is always the same: generate → evaluate → diagnose → improve → repeat.

How to Actually Build One

1. Pick one marketing task you do repeatedly. landing pages, ad copy, emails.. whatever you run most often.

2. Write down how you personally evaluate that output. what do you look for? what makes you say "this is good" vs "this is mid"? those criteria become your scoring checklist.

3. Turn each criterion into a pass/fail threshold. be specific. "thumb-stop power: 9/10 minimum" is a threshold. "make it good" is not.

4. Build the loop. tell claude: generate the output, score it against the checklist, diagnose any failures, rewrite, and re-score. do not stop until every criterion passes.

5. Add adversarial pressure. give claude a persona that attacks the output. a skeptical customer, a competitor, a distracted buyer. if the output survives the attack, it ships. if not, it iterates.

6. Save it as a reusable skill. now you're not prompting from scratch every time. you have a system that runs the same standard on every output.

i'm building a full library of these recursive skills for marketing teams. follow if you want them as they drop.