Anthropic Slashes Claude Code System Prompts by Over 80%

To adapt to new generation reasoning models like Claude Opus 5, Anthropic compressed Claude Code's system prompt from about 800 tokens to 164 tokens, a reduction of over 80%. The official team confirmed no measurable drop in coding test scores, noting that old "babysitting" prompts wasted tokens and caused instruction conflicts, while subtraction actually improved performance.

Confirmed

The team has removed many hard rules from the system prompt, leaving more judgments to the model. Anthropic also released a new context engineering guide on configuring CLAUDE.md and skills, and introduced the /doctor audit command.

Why it matters

This marks a shift in context engineering from adding rules to subtracting them. @jasonkneen noted that models may not need as many explicit instructions as imagined. However, as @caseyc2rd pointed out, this has sparked community debate over whether users should continue simplifying their local CLAUDE.md files or add more rules for smarter models.

2026-07-25 ~ 2026-07-26 · 11 related posts

Primary sources

1 near-duplicate retellings: AI寒武纪