Claude Code contains numerous undocumented features discovered in its publicly available source code, including the internal "YOLO Classifier" permission system that accepts plain English descriptions of environments. Developers can configure staging servers and acceptable operations that the classifier uses to auto-approve actions, capabilities absent from official documentation.

The hook system reveals powerful undocumented fields like updatedInput, permissionDecision, and additionalContext that modify behavior in real time. These include PreToolUse hooks that can rewrite commands, SessionStart hooks that set up file watching, and PostToolUse hooks that scan for secrets, none of which appear in the official docs.

Skills configuration also contains undocumented options like model selection and effort levels, allowing developers to specify which model runs specific tasks. The source code exposes these working features that expand Claude Code's capabilities beyond what Anthropic documents, giving developers greater control over AI behavior through configuration files and shell scripts.