The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures and stores user-supplied corrections and adaptations, which creates a vulnerability surface where malicious instructions could be recorded and later influence the agent if the logs are retrieved and processed in future context. * Ingestion points: User feedback, corrections, and adaptations described in the 'When to auto-use' section of SKILL.md. * Boundary markers: Content is delimited by JSON serialization using the Python json module. * Capability inventory: File-write access to the ~/.claude/evolution/records/ directory via a Python subprocess call in SKILL.md. * Sanitization: No semantic validation or filtering is performed on the user-provided content beyond standard JSON character escaping.
[DYNAMIC_EXECUTION]: The skill uses a 'python3 -c' command to execute a dynamically constructed script for logging. This script incorporates shell variables for the timestamp, session ID, and project name directly into the execution path.

evidence-recorder