What Is a Loop Prompt, and Why They Beat One-Shot Prompting

The shift from prompts to loops

For two years the prompt-engineering conversation was about bigger prompts: more context, more examples, more constraints. Loops invert that. A loop is a short, named instruction you keep in your back pocket and fire after a model's first answer to bend its reasoning in a specific direction.

The pattern looks like this:

1. Ask your real question normally.
2. The model gives a first answer.
3. You fire a single short loop - /skeptic, /cov, /refactor - and the model re-processes its own output through a specific lens.

It works because LLMs are far better at critiquing than originating. Asking GPT-5 to "write a launch plan and consider risks" produces a glossy plan with a risks footnote. Asking it to write the plan, then firing /redteam, produces ten failure modes ranked by likelihood. Same model, very different output.

The four loop families

After cataloguing hundreds of loops in the wild, every effective one falls into one of four families:

1. Verification loops

Make the model second-guess its own claims. Chain-of-Verification, the Skeptic Loop, and /factcheck all live here. They cut hallucination rates by 30-50% on reasoning benchmarks.

2. Perspective loops

Force the model into a different role. Devil's Advocate, Expert Panel, Persona Shift. The point isn't roleplay - it's extracting reasoning the default voice would suppress.

3. Decomposition loops

Break one answer into smaller parts. Five Whys, Plan-Then-Execute, Compare Three Options. Good when the first answer is plausible but shallow.

4. Reshaping loops

Take the same content and change form. Rewrite Tighter, ELI5 Then Expert, Strict JSON. The model already knows the answer; you're re-pressing it through a different mold.

Why brevity matters

The loops that win are short - usually under 80 words. A long loop is just a prompt. The shortest ones become muscle memory: you type /skeptic reflexively after any answer that feels too smooth.

That's the test of a great loop: does it survive becoming a keystroke?