Boris Cherny: Building Claude Code

Building agentic products requires an empirical, scientific mindset that prioritizes "unhobbling" models by removing rigid, outdated system prompts and scaffolding. As models like Opus 5 evolve, they frequently outperform the constraints imposed by legacy engineering practices. Developers should embrace an ablation-based approach, testing models with high-level, challenging tasks and allowing them to verify their own work through dynamic workflows and automated routines. This shift moves beyond traditional prompt engineering toward orchestrating thousands of agents to handle complex maintenance, such as refactoring large codebases or unifying abstractions. By treating models as organic, evolving entities rather than static systems, builders can unlock latent capabilities that significantly accelerate development, effectively automating tasks that previously required months of human labor. Boris Cherny, the creator of Claude Code, emphasizes that this iterative, data-driven process is essential for capturing the full potential of modern AI.