Deep Tide TechFlow message: On July 25, an Anthropic engineer disclosed that the team cut about 80% of Claude Code’s system prompt text for their latest Claude model, and that coding evaluations showed no performance decline. Earlier versions, due to insufficient model capabilities, had prompts containing many strict rules (such as prohibiting multi-line comments, etc.). But once the new model gained autonomous decision-making ability, these rules instead caused instruction conflicts and thinking to run off track. The new strategy changed to describing the goal (e.g., “write code that matches the surrounding code”) and dispersing instructions across tool descriptions and skill modules that are enabled only as needed. The team explicitly considers prompt bloat to be technical debt, and believes that as model capabilities improve, prompts should be simplified rather than stacked with more instructions.