ChainThink reports that on July 25, Anthropic engineers disclosed that the team reduced the system prompts for Claude Code by approximately 80% in the latest Claude model, with no performance degradation observed in coding evaluations.
Early versions included numerous strict rules in the prompts due to limited model capabilities, such as prohibiting multi-line comments; with the new model’s enhanced autonomous decision-making ability, such rules may now cause instruction conflicts and divergent thinking.
Anthropic has now shifted to describing objectives, such as “write code that matches the surrounding code,” and has distributed instructions across tool descriptions and skill modules enabled on demand.
The team views prompt inflation as technical debt and believes that once model capabilities improve, instructions should be simplified rather than continually added.
