Your Old Prompts Are Making Claude Worse - Lance Martin, Anthropic

Frontier models often operate below their potential due to minor configuration errors and outdated prompting techniques. Auditing prompts to remove unnecessary "thoughtfulness boosters" and "verification rituals" improves performance while reducing costs. Implementing prompt caching significantly lowers expenses for multi-turn applications, as cached tokens are billed at a fraction of base rates. Effective evaluation design—focusing on headroom, low variance, and clear rubrics—enables autonomous "hill climbing," where models iteratively improve their own prompts and skills. For long-duration tasks, decoupling the agent's "brain" (harness) from its "hands" (sandbox) ensures reliability and security. By utilizing these primitives, developers can transition from simple chat interfaces to sophisticated, autonomous agents capable of executing complex, multi-day workflows while maintaining cost-efficiency and performance.