STUFF: The fastest way to unstick an LLM is to stop talking to it.

STUFF: The fastest way to unstick an LLM is to stop talking to it.

When a model snarls up at the edge of its knowledge, more prompting rarely helps. It's locally consistent but globally wrong, and every extra turn deepens the rut.

What works: an external check. Hand the output to a different process — another LLM with a narrower job, a type checker, a test, a query that has to actually run. Then feed the result back.

The trick isn't the second model being smarter. It's that the signal comes from outside the generator's own distribution. New provenance breaks the autocorrelation. It's noise that lets the system escape a local minimum.
I've been running this pattern on real blockers. Things that looked like dead ends resolve in one or two passes (of course not always).

Two things to watch: — If your checker shares the generator's blind spot, they'll confidently agree on something wrong. Use a different model, or better, a tool that actually executes. — Vague feedback creates oscillation. The model just rewords. Structured, specific feedback ("this contradicts X", "this step has no justification") is what moves it.

Generator-verifier loops aren't new. But most people still try to fix stuck LLMs by prompting harder. The leverage is in the loop, not the prompt.

So why do this? I started to do it to use 2 X LLMs where one pretends to be human in loop and vice versa so I could just sit back, watch and eat popcorn... Now I am thinking this is a far better way of doing decision support...

scottg/out