gauntlet-looplisted

# Gauntlet Loop Source: https://somethingbig.ai/gauntlet-loop (Matt Shumer). One idea. Pick a real example of great, then have a builder and a separate critic argue about the gap until the gap closes. This skill spawns subagents. Invoking it is the request for them. ## 1. Set the bar first The bar is the whole method. It must be something an agent can actually open and compare against: screenshots, a live URL, a reference file, a writing sample, a recorded metric. Not a bar: "make it amazing", "production-ready", "keep improving it". Those give the critic nothing to measure, so the loop degenerates into praise. Ask the user for the reference if they did not give one. Offer candidates from the domain. Confirm the bar in one line before spawning anything: ``` bar: [what the critic compares against, and how it inspects it] ``` If the reference is a file or URL, verify it is reachable now. A critic that cannot open the bar will invent one. ## 2. Decompose Break the goal into the smallest pieces that can be improved independently. Independent means two builders can work at once without editing the same thing. Aim for 3 to 8 pieces on a first run. State the pieces before starting. The user's correction here is cheaper than five rounds of the wrong decomposition. ## 3. Loop each piece Per piece, per round: 1. **Builder** makes or revises the artifact. It does not judge its own work. 2. **Critic** runs in a fresh agent with no builder history and no builder rational