Loop Until Done

This skill forces the AI to work in a self-evaluating iterative loop until the output meets a strict quality bar. Instead of producing a single-shot answer, the agent plans, executes, scores its own work against explicit criteria, and iterates on the weakest point until every criterion scores 8 or higher.

Inspired by Andrej Karpathy's AutoResearch loop and the "Bilevel Autoresearch" paper — applying the principle that repetition with honest self-evaluation produces dramatically better results than single-pass generation.

When to Use This Skill

Activate when the user:

• Says "loop until done", "iterate until perfect", "keep improving until it's good"
• Asks for high-quality output that must meet specific criteria
• Wants exhaustive refinement without manual back-and-forth
• Needs production-grade deliverables (documentation, code, architecture, specs)
• Says "don't stop until it's right" or "be strict with yourself"

How It Works

The agent enters a structured loop with four phases per iteration: