The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using grep to scan through session logs for specific patterns of user feedback. It also uses ls to check for the existence of local overlay directories specified by environment variables.
[DATA_EXFILTRATION]: The skill accesses and reads sensitive local files, including historical session transcripts (~/.claude/projects/*/*.jsonl) and project-level configuration files (CLAUDE.md). This data is analyzed locally to derive filtering rules, but no evidence of network transmission was found.
[DYNAMIC_EXECUTION]: Based on its analysis of user transcripts and interview responses, the skill generates and writes a JavaScript file (~/.claude/vestige-patterns.js) containing regex-based filtering logic. This involves the creation of executable logic at runtime from processed data. Additionally, the skill supports an overlay mechanism that loads and follows instructions from a directory specified by the $POWERSUIT_OVERLAY environment variable.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical data, creating an attack surface where past messages could influence future behavior.
Ingestion points: Reads from historical session logs (~/.claude/projects/*/*.jsonl) and local configuration files (CLAUDE.md).
Boundary markers: No explicit delimiters or instructions are used to distinguish ingested log data from the skill's own logic during the analysis phase.
Capability inventory: The skill possesses the ability to perform shell commands, read local files, and write new skill definitions and JavaScript code.
Sanitization: There is no mention of sanitizing or escaping the content retrieved from transcripts before using it to generate new filtering rules.
[PERSISTENCE]: The skill creates persistent artifacts in the user's home directory (~/.claude/skills/ and ~/.claude/vestige-patterns.js). These files remain active in the environment, modifying the agent's output behavior across subsequent sessions.

calibrate