How this channel uses it In the captured material, a judge prompt is the tool by which taste gets engineered rather than left to chance: instead of relying on a single model's default behavior, the team writes a dedicated prompt that tells an LLM judge what good looks like — encoding the character's voice, tone, and quality bar into an explicit evaluation prompt [^259b88ca-e6ae-534f-b08e-5773fe2ab869@133]. This is how Portola kept Tolan — an embodied AI alien companion — feeling consistent and on-character even as output was generated by improvisation rather than a script [^259b88ca-e6ae-534f-b08e-5773fe2ab869@197]. Why it matters here Tolan's quality bar was defined by taste, and the judge prompt is the mechanism that turned that taste into a repeatable check on model output [^259b88ca-e6ae-534f-b08e-5773fe2ab869@100]. - A judge prompt tells an LLM judge what to evaluate and by what standard.- In Tolan's build, it is described as the way taste gets engineered into LLM judges.- It complements character-driven output: the judge prompt checks that improvised responses stay on-character.