Claude Code adds eval design and hillclimb commands

Claude Code adds build-eval and hillclimb commands for AI evaluation and optimization, addressing self-deception, with criteria and anti-overfitting protocol. Case studies show improved accuracy and ~5x cost reduction.