Skip to main content
AI RADARModels
268•07/25/2026, 08:15•2 min read

New Context Engineering Rules for Claude 5 Models

New Context Engineering Rules for Claude 5 Models

Tariq from Anthropic has published new context engineering rules for Claude 5 models.

These principles have also allowed for a reduction of the system prompt for Claude Code by approximately 80% without measurable degradation in results on their coding benchmark.

tldr:

• CLAUDE.md is only for pitfalls. Remove everything that the model can already understand by studying the repository, and spend tokens on what it can actually stumble over. • Replace hard prohibitions with style descriptions. Previously, the rule was stated as never write comments. Now it is write code that matches the density of comments and naming in the surrounding code. • Do not duplicate the same rule in the system prompt and tool description. Keep it only in the tool description. • Break long skills into several files and upload only what is needed for a specific task. • The specification does not have to be in Markdown. A test set, HTML layout, or function from another codebase will suffice.

This is more useful than the next ten skills.

Read his material today, and then run /doctor to shorten your own CLAUDE.md and skills.

Why it matters

Analysis

These new rules can significantly enhance the efficiency of working with AI models, reducing resource expenditure and improving outcomes. They open up new opportunities for developers in utilizing AI technologies.

Discuss in community

Share your questions and insights with developers

Join Telegram
+5 Points
← Back to news feed

Related in Ecosystem

Practical guides, core terminology, and agent skills