We Cut 80% of Claude Code’s Prompt

Cherny says each model learns different capabilities, sometimes in unexpected ways. He describes Opus 5 as able to run for days or weeks on tasks and says its resistance to prompt injection combines model alignment, an injection classifier, and an auto mode classifier.

The Claude Code team removed more than 80% of its system prompt because Opus 5 handles behaviors earlier models needed instructions for. Cherny recommends ablation: remove prompts and tools, then add instructions back only when real use shows the model repeatedly needs them. He says evals can also become saturated after one to three model generations.

Cherny describes product overhang as capability that products fail to elicit, or product design that gets in the model’s way. He recommends giving models harder, higher-level tasks with clear guardrails and exit criteria. As examples, he describes a Claude Code workflow that rewrote more than 100,000 lines from Zig to Rust over 11 days, and a Swift rewrite of Claude’s desktop app that was still running after two weeks with screenshot comparisons for verification.

Cherny says dynamic workflows split a complex task into stages, using agents in sequence and in parallel; loops and cloud routines repeat tasks on a schedule. Anthropic uses routines to find dead code, add or remove tests, ship completed experiments, and unify duplicated abstractions across codebases. He says roughly 20 to 30 routines run daily, sometimes coordinating hundreds or thousands of agents.

Cherny qualifies that coding is solved for some of his work, while deep systems code, distributed systems, and precise UI verification remain difficult. He says strong users test the model, observe where it struggles, and adjust rather than relying on assumptions from earlier models. For students, he recommends combining computer science with practical skills such as product design, business, data science, and talking to users.