Anthropic Cut 80% of System Prompts, Shifting to Minimalist Context Engineering

机器之心 · wechat · 2026-07-25

Anthropic's technical team found that with the release of new reasoning models like Claude Opus 5, old "babysitting" prompt engineering not only wastes tokens but can cause conflicting instructions and degrade performance. They deleted over 80% of the Claude Code system prompts without observing any measurable performance loss.

The team summarized new paradigms for Context Engineering:

From Rules to Judgment: Rigid guardrails (e.g., defaulting to no comments) were removed; new models can make correct decisions based on surrounding context.
From Examples to Interfaces: Instead of piling examples, focus on designing expressive tool parameters and enums to guide the model.
Progressive Disclosure: Split low-frequency but crucial instructions into independent Skills or lazily-loaded tools to keep the main context uncluttered.
Auto Memory & Rich References: Models can now save memories automatically and use HTML artifacts or codebases as high-fidelity reference materials.

Developers are advised to keep CLAUDE.md strictly lean (under 60 lines for high-performing teams) and can use the /doctor command in Claude Code to automatically trim legacy projects.

Original post →

Anthropic Cut 80% of System Prompts, Shifting to Minimalist Context Engineering

More from coding & agent