Claude Code's system prompt shrank by over 80%, scores held
Over 80% of Claude Code's system prompt is gone for Claude Opus 5 and Claude Fable 5, and Anthropic says its coding evaluations showed no measurable loss. Anthropic's new context engineering guide explains why: the old rules overconstrained the model.
Internal transcripts showed a single request carrying both "leave documentation as appropriate" and "DO NOT add comments." The new prompt tells Claude to match the surrounding code's comment density, naming, and idiom.
Progressive disclosure works like a reference shelf: Claude pulls a skill or tool definition only when the task needs it. The guide suggests a short CLAUDE.md focused on codebase gotchas, and /doctor in Claude Code rightsizes skills and CLAUDE.md files.
One catch: the post doesn't name the coding evaluations behind the "no measurable loss" claim.
Related stories
- Anthropic tells Opus 5.5 users to delete "think hard" lines
- Prompt audit in Claude Code cuts cost 14.6%
- 80% of Claude Code's system prompt got cut
- Claude's hillclimb cut support costs to about a fifth
- Claude Code sessions keep running after you close the laptop
- Claude Code skips AGENTS.md when telemetry is off
Comments
No comments yet. Be the first.
Join the conversation
Sign in with Google to leave a comment. Your name and avatar come from your Google profile, and the comment appears after moderation.