Anthropic removed over 80% of the Claude Code system prompt for Claude 5 generation models (Opus 5 and Fable 5), with no measurable losses on coding evaluations. The July 24 post explains the shift in approach: more capable models need fewer rigid rules and handle context better by relying on their own judgment.

If you build agents or use Claude Code with CLAUDE.md files and Skills loaded with instructions, this changes how you work. The old pattern was to accumulate defensive rules to avoid worst-case scenarios: “don’t write comments”, “don’t create planning documents”, “don’t delete files”. With Claude 5, many of these instructions end up contradicting each other and force the model to compromise before deciding. Removing them frees it instead of weakening it.

Anthropic has made the /doctor command available in Claude Code to review and resize Skills and CLAUDE.md files. We haven’t tested it yet.

The paradigm shift is straightforward: you used to give rules, now you ask for judgment. Rules worked when the model could cause serious damage on its own initiative. Now that the model is more capable, over-constraining costs more than the risk it prevents.

Alongside the Agentic patterns Cookbook published the same day, Anthropic is making public its internal practices on how to work with Claude in production: one on architectural patterns, the other on how to write context.

In detail

What context engineering means.

When you write a prompt for Claude, what you send is only part of what the model receives. The rest comes from the system prompt (base instructions), CLAUDE.md files (project rules), Skills (specific competencies), memory and other sources. Anthropic calls all this context engineering: the practice of assembling the right context so the model works well across many different requests, not just one.

The difference between prompt and context is that context needs to be general. You don’t know what the user will ask, so you can’t be too specific. But if you’re too vague, the model goes in directions you don’t want. Finding the right balance is hard, and it changes when the model changes.

What changed with Claude 5.

Anthropic found it had over-constructed Claude Code. Reading transcripts of internal sessions, it found conflicting instructions in the same message: the system prompt said one thing, Skills said another, and the user request a third. For example, one instruction said “leave documentation where appropriate” while another imposed “DO NOT add comments”.

The model had to compromise between these conflicting directions before even starting work. With Claude 4, these rules helped avoid worse behaviors. With Opus 5 and Fable 5, the model has enough judgment to decide on its own when to leave a comment and when not to. The confirmation is in the numbers: remove over 80% of the system prompt, coding evaluations didn’t deteriorate in measurable ways.

There’s also an instrumental factor. Before, CLAUDE.md served as memory, information source and guide. Now Claude Code has separate tools for each function: persistent memory, artifacts and Skills. Context no longer has to do all the work alone, and instructions that duplicated these functions become dead weight.

The limits of what we know.

The 80% figure is stated by Anthropic on its own internal evaluations. We don’t know which tests it used or how rigorous they were. The advice applies to Claude 5 models: on less capable models, removing defensive rules might degrade performance. The shift from “rules” to “judgment” also requires trusting the model on decisions where you previously had explicit control. It’s not a jump you make everywhere the same way: on a system that handles sensitive data or performs irreversible operations, initial caution remains reasonable even with a capable model.