The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is specifically instructed to access and analyze the contents of the ~/.claude/ directory, including session history, rules, and project memory files. These directories are internal to the AI agent's environment and typically contain sensitive conversation context, private project data, and application configurations.
[COMMAND_EXECUTION]: The skill performs extensive file system operations using Read, Glob, and Grep on absolute paths (e.g., /Users/gabe/claude-skills/) and internal hidden directories. This behavior grants the agent visibility into the local file system beyond the immediate workspace.
[PROMPT_INJECTION]: The skill relies on 'Deep Context Analysis' by ingesting untrusted data from previous conversation histories and external session files (session-context.md). This historical data can contain malicious instructions from previous sessions that could influence the agent's current orchestration logic or skill recommendations through indirect prompt injection.
Ingestion points: Reads full conversation history and internal memory files from ~/.claude/.
Boundary markers: None provided to distinguish between historical instructions and data.
Capability inventory: Can execute network requests (WebFetch), file reads, and search operations based on processed history.
Sanitization: No evidence of input validation or instruction filtering for historical data.
[SAFE]: The skill includes a logging mechanism to track performance and recommendations. While this involves writing a persistent state file (scout-pro-usage-log.json), the activity is aligned with its stated purpose of improving recommendations over time.

scout-pro