1One-shot prompting leaves quality on the table
The first draft an AI model gives you is rarely its best work — it is its most average work. The model answers immediately, with the highest-probability words, not the best-judgment words. Iteration is how you push past the average: each pass with specific feedback moves the output toward what you actually wanted.
Loop generation formalizes this. Instead of you manually copying the output back in with "make it better," the loop takes the draft, critiques it against your goal, and rewrites it — for as many rounds as you set.
2How a loop actually works
A loop has three steps: generate a first draft, critique the draft against your original goal and constraints, and rewrite incorporating the critique. Round two critiques the round-one rewrite, and so on.
The magic is in the critique step, not the rewrite. A good critique names the specific weakness — "the hook is generic, the middle loses specificity, the call to action is weak" — so each rewrite is a genuine improvement instead of a shuffle of the same words.
3Why more loops are not always better
Improvement is not linear. The first loop usually delivers the biggest jump; loops two and three tighten things; by loop five you are often just moving furniture. Diminishing returns hit hard, which is why most real workflows settle at 2-3 loops — enough to lift quality, few enough to keep latency and cost sane.
4Looping for prompts vs. looping for content
There are two very different use cases. Looping a prompt refines the instructions themselves until the prompt is bulletproof — you invest once and reuse forever. Looping content refines a single piece (an email, a script, a landing page) to publish quality.
For prompts, the loop should critique clarity, completeness, and ambiguity. For content, it should critique persuasion, specificity, and flow. Using the right critique lens for the right loop is the difference between polishing and polishing a turd.
5Getting started without burning credits
Start with loops of 2 on low-stakes drafts until you feel the rhythm. Keep your original goal visible to the loop — a loop with no destination wanders. And always keep the final draft of a good loop run; a great looped prompt is worth saving to your library, because the next time you need that output, the first draft already starts above average.