Uncovering hidden capabilities in Claude Code

Analysis of Claude Code version 2.1.87’s source code reveals several undocumented features. A key finding is the YOLO Classifier permission system. It allows describing environments in plain text so the AI understands the security level of operations.

Specialized agents can be built using persistent memory at the user and project levels. A self-learning system was also found: the autoMemory and autoDream functions enable the model to extract knowledge from sessions and consolidate it every 24 hours. This creates a continuous learning loop without overtraining the model.