Since the recent Claude updates, the model has started describing every single step in detail before taking action, producing huge blocks of explanatory text even for very simple tasks.
This causes tokens to disappear rapidly—like a tap left open all day. The verbose narration of thoughts and tool calls defeats the purpose of token efficiency and caveman-style compression.
Here are examples of the narration bloat observed during runs (private details redacted with blur):
Please look into ways to aggressively compress or suppress this verbose prose, especially in tool call steps, to avoid wasting thousands of tokens.
Since the recent Claude updates, the model has started describing every single step in detail before taking action, producing huge blocks of explanatory text even for very simple tasks.
This causes tokens to disappear rapidly—like a tap left open all day. The verbose narration of thoughts and tool calls defeats the purpose of token efficiency and caveman-style compression.
Here are examples of the narration bloat observed during runs (private details redacted with blur):
Please look into ways to aggressively compress or suppress this verbose prose, especially in tool call steps, to avoid wasting thousands of tokens.