Fetch and execute mizchi's empirical-prompt-tuning skill at runtime. Use when evaluating or iteratively refining an agent-facing prompt (skill / slash command / task prompt / CLAUDE.md section / code-gen prompt) by dispatching an unbiased subagent, then improving until metrics plateau. Trigger right after creating or heavily revising such a prompt, or when agent misbehavior is suspected to stem from ambiguity in the instruction.