Skip to content- Sources: primary
- Summary: Lance Martin presents bundled Claude Code skills that strip outdated prompt patterns such as think step by step and contradictory rules, check prompt cache hit rates, build evals and hill climb on them with a train and test split to catch overfitting. He reports that cache fixes on internal benchmark harnesses cut cost by up to 72 percent with no meaningful quality change. A cost-focused hill climb on a support ticket eval moved from Opus 4.8 high effort to stronger models at low effort for about 90 percent lower cost and better scores. He also advises binary or pairwise LLM judge rubrics checked across several models and replicate runs.
- Why it matters: Teams on the Claude API can run these audits on existing code and test a stronger model at low effort before buying more tokens.