Building on LLMs: delete your system prompt, let the model run

Boris Cherny reveals that building on Claude Code means constantly deleting and rebuilding prompts. For Opus 5, they deleted 80% of the system prompt because the model no longer needed it. The key insight is product overhang: models can do far more than current products allow. Teams should give high-level tasks with verification tools and

let the model run for days or weeks. The real skill is not prompt engineering but elicitation—figuring out what the model can already do. Dynamic workflows let Claude orchestrate thousands of agents for complex rewrites, like the Bun runtime from Zig to Rust in 11 days. Self-maintaining routines now clean up dead code, unify

abstractions, and write tests daily. Coding is solved for many common cases, but deep systems and distributed code still struggle. The biggest mistake is over-specifying instructions. Instead, treat the model like a smart coworker: give it a goal and a way to verify its own work. The empirical approach—try, observe, adjust—beats any theoretical framework.