Why Anthropic Cut 80% of Claude Code System Prompts Overnight

When Claude Opus 5 was launched, Anthropic’s team announced that the new, stronger reasoning model rendered many of the old "babysitter" style system prompts obsolete. They removed over 80% of the Claude Code system prompts and observed no measurable degradation in coding evaluation performance.

Key insight : the more capable the model, the less it needs verbose, hand‑crafted rules. Long prompts waste tokens and can even hurt results. New guidelines recommend keeping Claude.md under 60 lines (never exceeding 300) and delegating task‑specific context to Skills .

Context engineering is defined as the combination of system prompts, Skills, CLAUDE.md, memory, and other sources that together shape the model’s behavior. Unlike static prompts, context is reused across requests and must be generic enough to handle unknown user inputs.

Anthropic illustrated the shift with a concrete example of contradictory instructions that appeared in the same request, forcing Claude to repeatedly weigh conflicting rules. They concluded that many of these constraints were historically necessary to avoid worst‑case failures, but can now be dropped in favor of letting the model rely on surrounding context and its own judgment.

Old system prompt example: "When writing code, never write comments. Never use multi‑line docstrings – at most one short comment line. Unless the user explicitly asks, do not create planning, decision, or analysis documents – work based on dialogue context, not intermediate files."

The team replaced it with a concise rule that tells Claude to match the surrounding code’s comment density, naming style, and idioms, allowing the model to decide when comments are appropriate.

Other best‑practice shifts include:

From examples to interface design : instead of providing many usage examples that constrain the model, design clear tool interfaces (e.g., a Todo tool with three enumerated states) so Claude can infer behavior.

From monolithic prompts to progressive disclosure : split large instructions into multiple Skills or files that are loaded on demand, reducing token waste.

From repeated instructions to concise tool descriptions : move repeated usage notes into the tool’s own description rather than the system prompt.

From manual memory entries to automatic memory : Claude now automatically saves relevant context instead of requiring users to insert # shortcuts.

From simple markdown specs to rich reference materials : Claude can now ingest HTML artifacts, test suites, or full codebases as reference material, enabling more sophisticated verification.

Anthropic packaged these practices into the /doctor command (Claude doctor), which automatically trims system prompts, Skills, and CLAUDE.md to an optimal size.

For developers building their own agents, the article advises focusing effort on the system prompt (tightly bound to product context), keeping CLAUDE.md lightweight, using Skills as lightweight guides, and treating reference materials as explicit files referenced with @ mentions.