Developer Builds Self-Improving AI Prompt System With 6-Check Gate After 4,150 LLM Calls

A developer built a self-optimizing AI prompt system in which a large language model proposes edits to its own prompts but holds zero authority over whether those edits are accepted. Every proposal is evaluated through an A/B testing engine and a six-step deterministic gate covering checks such as statistical confidence, edit distance, and content drift before any change is promoted. The entire system ran locally on a MacBook using a Qwen 4B model via MLX, completing 4,150 LLM calls at no cost and finishing each iteration in roughly 42 seconds. A critical bug discovered during development had the confidence check inverted, allowing near-random results to pass, which masked further bugs for weeks until corrected. The project's key finding is that the safety gate — not the optimizer — is the most important component of any self-improving AI system.