Why Deleting 80% of Claude Code System Prompts Doesn’t Hurt Performance

Introduction

Thariq, a founding member of Anthropic’s Claude Code team and lead of the Claude Agent SDK, published a detailed analysis revealing that deleting over 80% of the system prompts in Claude Code caused no detectable degradation in coding evaluation metrics.

What Context Engineering Means

Context engineering encompasses everything the model receives for each request: system prompts, Skills, the CLAUDE.md file, memory, and other sources. The user’s message is only a small part; the rest is automatically assembled by the system.

Why Rules Were Needed Before

Earlier models lacked sufficient judgment. Hard rules such as “never write comments” or “always keep documentation” prevented catastrophic failures like stray comments or accidental file deletions. These constraints often conflicted—system prompts might say “preserve documentation,” a skill could forbid comments, and a user instruction might add a third demand—forcing the model to resolve contradictions.

Model Strength Enables Rule Removal

With Claude Opus 5 and Claude Fable 5, the model’s judgment is strong enough to handle decisions without explicit safeguards. The team removed more than 80% of the system prompts and observed no loss in coding benchmarks. The former rule “never write comments” was replaced by a softer intent: “produce code whose comment density, naming, and style match the surrounding code.”

Six Revised Best Practices

From Rules to Judgment: Replace rigid constraints with high‑level intents, allowing the model’s reasoning to choose appropriate behavior.

From Examples to Interface Design: Instead of feeding example snippets, define clear tool parameters (e.g., an enum of task states) that guide the model.

From Full Pre‑loading to Progressive Disclosure: Split context into a file tree and load only the needed parts on demand; rarely used skills become separate files.

From Repetitive Instructions to Concise Tool Descriptions: Embed usage details directly in the tool description, eliminating duplicated instructions in the system prompt.

From Manual Memory in CLAUDE.md to Automatic Memory: The model now automatically saves relevant context, removing the need for manual # shortcuts.

From Simple Specs to Rich References: Use code files, HTML artifacts, test suites, or functions from other repositories as high‑fidelity specifications rather than plain markdown plans.

Applying the Changes to Your CLAUDE.md

Keep CLAUDE.md lightweight: a brief description of the repository’s purpose, with detailed constraints moved to Skills. Use @ mentions to reference code files, and only retain hard rules for high‑risk scenarios where mistakes have severe consequences.

Anthropic also provides a /doctor command in Claude Code to help trim Skills and CLAUDE.md to an appropriate size.

Conclusion

As large‑language models become more capable, the prompt‑engineering paradigm must evolve from rule‑heavy, example‑driven contexts to leaner, intent‑driven configurations that let the model exercise its own judgment.