XNews

Claude Code System Prompt Cut 80%: Fewer Examples, More Model Judgment

Claude Code System Prompt Cut 80%: Fewer Examples, More Model Judgment

Anthropic's Claude Code team revealed they cut the system prompt by 80%, finding that removing few-shot examples improves performance because modern models are more creative than given examples. They shifted from hard prohibitions to providing context, and ensure every instruction is 100% correct with no exceptions. This approach relies on frontier model judgment. They also discussed Claude Tag (65% of PRs now AI-landed), automated code review with trust built over six months, and evaluation as the foundation. Tool design trends toward fewer, non-overlapping tools. The team emphasized that human value shifts to product taste and ambition, not execution.

Source posts