Anthropic's Boris Cherny on Cutting 80% of Claude Code's System Prompt
Boris Cherny, creator of Claude Code at Anthropic, discusses Opus 5 and the philosophy behind building the harness. He explains that Anthropic deleted over 80% of Claude Code's system prompt for Opus 5 because the model is intelligent enough to no longer need most instructions, and recommends an ablation-based approach: delete the prompt, use the product, observe failures, then add only the lines the model genuinely needs. Cherny frames prompt engineering as shifting from line-by-line optimization to model elicitation, introducing the concepts of "product overhang" (capabilities the model already has but products don't expose) and "hobbling" (harness scaffolding that gets in the model's way). He argues that best practices shift every model generation, so builders must be empirical and willing to "press delete" on accumulated prompts, skills, and hooks roughly every six months. Evals are more stable but still saturate and need refreshing. The key skill is giving the model hard tasks, letting it verify its own work, and avoiding over-specification. He illustrates this with examples: Claude rewriting the Bun runtime from Zig to Rust in 11 days using dynamic workflows, and a two-week task rebuilding the Claude desktop app from Electron to Swift with pixel-level verification. Cherny also describes new orchestration primitives—dynamic workflows for productive multi-agent tasks, plus loops and routines for repetitive cron-style maintenance. Anthropic now runs dozens of daily routines (dead code cleanup, test coverage, "abstraction police") across its codebases, with thousands of agents executing continuously. On advice for builders, he says coding is "solved for the kind of coding I do" but still hard in deep systems code, distributed systems, and pixel-perfect UI, and recommends learning computer science through building practical things you want, combining engineering with product, design, and business skills.
Original sources
YouTube Y Combinator
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.