5 Ways a Gauntlet Loop Breaks (And How to Fix Each)

5 Ways a Gauntlet Loop Breaks (And How to Fix Each)

The gauntlet loop looks robust from the outside: builders, critics, a bar, iteration. In practice most first runs fail, and they fail in predictable ways. Here are the five failure modes we see, worst first.

1. The vague bar

The most common killer. If the bar is "award-winning design" or "world-class writing", the critic has nothing to fetch, so it invents a comparison and approves everything. The loop spins happily and produces mediocrity with confidence. Fix: the bar must be named (a specific page, post, repo), fetchable (the critic can open or screenshot it) and comparable (both artifacts can sit side by side). If you cannot paste a URL to your bar, you do not have one yet.

2. The builder grading itself

If the same agent context builds and judges, the judgment inherits every bias: effort, attachment, fatigue. Fix: the critic is a separate sub-agent with fresh context. It sees two artifacts and nothing else - not the plan, not the round count, not the excuses.

3. The soft critic

A critic asked "how good is this, 1-10?" gives 7, then 8, then 9. Scores drift because they are judged against expectations, and expectations move. Fix: give the critic a binary job. Which one is better, A or B, labels stripped? And require it to name the single biggest gap when the bar wins - that gap is the next round's entire instruction.

4. The fixed round count

"Improve it 5 times" stops exactly when the model feels like coasting, whether the work is great or garbage. Fix: the exit is winning the blind comparison, or a human calling it. Nothing else ends the loop.

5. Pieces too big to judge

A critic can compare a hero section, a weapon model, one chart. It cannot honestly compare "the whole app" in one look, so it rounds up. Fix: split the goal into artifacts small enough that a blind pick is meaningful, and gauntlet each piece.

The pattern behind the fixes

Every failure is the same failure: somewhere, self-judgment leaked back in. The template encodes all five fixes; the annotated original shows them in Shumer's own words.