CloudCodeTree LogoCloudCodeTree
AI NewsTutorialsAbout
CloudCodeTree Logo
CloudCodeTree
  • AI News
  • Tutorials
  • About
← Back to AI News
Anthropic Cut 80% of Claude Code's System Prompt for Opus 5 — Here Are the Six Rules They Found

Anthropic Cut 80% of Claude Code's System Prompt for Opus 5 — Here Are the Six Rules They Found

Chris Harper

2 min read

Jul 29, 2026 · 12:06 UTC

AI
Workflow
Claude Code
Best Practices

Anthropic's Claude Code team removed 80% of their system prompt when moving to Opus 5 and Fable 5 — with no measurable regression on coding evals — and published the six context-engineering shifts that made it work.

The new rules of context engineering for Claude 5 generation models (Thariq Shihipar, July 24) is one of the most practically useful pieces Anthropic has published for developers who write CLAUDE.md files or build agents. The headline finding: the rules-heavy prompts that shaped Claude 3.x behavior actively impair Claude 5 models. The model is smarter — over-specifying gets in its way.

The six shifts

1. Rules → judgment. Replace "never do X" with the reasoning. Old: "default to writing no comments. Never write multi-paragraph docstrings." New: "write code that reads like the surrounding code: match its comment density, naming, and idiom." Opus 5 infers the constraint from the reasoning.

2. Examples → interface design. Move few-shot examples out of the system prompt and into tool/schema definitions. Claude 5 models read structured interfaces more reliably than inline examples.

3. Upfront context → progressive disclosure. Don't front-load all project context. Inject it when it becomes relevant — the model tracks what it needs and when.

4. Repetition → one clear statement. Saying the same constraint three times reinforces it with older models; it introduces noise for Claude 5. Each rule: state once, clearly.

5. Manual memory → auto-memory. Claude 5 models learn what to remember from interaction patterns. Explicit "always remember X" instructions are largely redundant and add clutter.

6. Simple specs → rich references. Point to actual files rather than summarizing them. See tsconfig.json for our TS settings outperforms a paragraph describing those settings. The model reads the file.

How to apply to your CLAUDE.md right now

  1. For every "always/never" rule, ask: does the model already understand why? If yes, replace the rule with the reasoning or delete it entirely.
  2. Replace any paragraph describing a convention with a path to the file that embodies it.
  3. Remove duplicated constraints — if the same idea appears twice, keep one.
  4. Move step-by-step examples into Skills (.claude/skills/), not the main CLAUDE.md.

Sources: The new rules of context engineering for Claude 5 generation models — claude.com · Developer analysis — developersdigest.tech · Anthropic 80% prompt removal — AI Weekly