robonuggets/gauntlet-loop

What it solves

It prevents AI agents from stopping at "good enough" by replacing vague rubrics with a concrete, external quality bar. It automates the creation of a rigorous prompt that forces an agent to loop through building and criticizing until the output is objectively better than a specific, high-quality reference.

How it works

1. Goal Setting: The user provides a goal (e.g., a landing page or a CLI tool).
2. Bar Selection: The skill suggests 2-3 specific, fetchable, and comparable real-world examples (the "bars") to serve as the quality standard.
3. Prompt Generation: Once a bar is selected, the skill generates a short, paste-ready prompt (approx. 150 words).
4. Execution Loop: When pasted into an agent session, the agent splits the work into small pieces and employs a "builder" and a separate "harsh critic" agent. The critic performs blind comparisons between the current work and the reference bar, looping until the work wins the comparison.

Who it’s for

Developers and power users of AI agents who want to ensure their agentic workflows produce professional-grade results rather than mediocre outputs.

Highlights

• Blind Comparison: Uses a separate critic agent with fresh context to pick a winner between the work and the bar, avoiding the score-drift common in numeric grading.
• Concrete Standards: Rejects vague bars in favor of named, fetchable, and comparable references.
• Agent Agnostic: Works with Claude Code features like /loop and ultracode, but can be adapted for any agent via plain instructions.
• Outcome-Based Exit: The loop only ends when the work wins the blind comparison or the user manually stops it.