22 August 2026

Claude's system prompt shrunk 80% as model absorbed harness functions

• Anthropic reduced Claude Code's system prompt by 80% after the model learned to handle tasks internally through reinforcement learning in real environments.
• Models are absorbing capabilities that previously required external instructions, like managing context window size and calling tools, into their learned weights.
• As models internalize these functions, engineers can remove the scaffolding code that once directed behavior, simplifying the overall system architecture.

How it was covered

Latent Spaceswyx & Alessio

Through reinforcement learning trained in real-world environments, models are beginning to internalize harness capabilities like context window management and tool calling. As models absorb these functions, engineers can delete the external harness scaffolding. Anthropic recently deleted 80% of Claude Code's system prompt as the model absorbed those capabilities.