- Anthropic removed ~80% of Claude Code's system prompt for newer models, finding less direction needed as models become smarter.
- Experiments with a small, fast, cheap model (deepseek-v4-flash) showed cutting prompts in half (~5.0k to ~2.3k chars) yielded no performance drop.
- Results: mean reward slightly higher (0.596 vs 0.573) with short prompts, and median input tokens decreased 32% in same-outcome tasks.
- No capability cliff was observed; task flip rates were consistent with single-attempt variance, and no task category collapsed.
- The takeaway is that accumulated prompt text may silently cost money without helping, and new model generations should prompt deletion rather than addition.