Claude Code gauntlet loop: subagent prompting tutorial

Jay explains the gauntlet loop as a prompt structure for an AI agent: define the task, specify how subagents should build and review it, and set a quality threshold for stopping. Workers handle individual parts while separate critics evaluate their output. He connects this approach to the evaluator pattern discussed in Anthropic's 2024 article on effective agents and presents Claude Code as a way to try it through prompting.

The first test uses a floor plan and reference photos from an apartment listing in Darling Point, Sydney. Jay identifies the model as Opus 5 and asks it to build an explorable 3D walkthrough. After roughly two hours, the workflow is still running. He shows a report comparing reference photos with generated screenshots, including rounds the critics marked as failed. The walkthrough resembles the supplied layout in his assessment, though he points out unfinished furniture and textures.

A second test produces a Ketone IQ landing page after about an hour and 19 minutes, with light and dark modes and animations. Jay notes that its design differs from the brand's actual website. His recommendation is to establish an MVP or design system before using the loop to refine it, since repeated review can consume substantial time and tokens while pursuing the wrong direction. He also introduces /gauntlet-loop, a skill that generates this prompt structure from a task.