The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive local application directories to read conversation transcripts and session history from various coding agents (Claude Code, Cursor, Codex, OpenCode, and Pi). Accessed paths include ~/.claude/projects/, ~/.cursor/prompt_history.json, ~/.codex/history.jsonl, and others in the user's home directory. This represents a significant exposure of private user interaction data.\n- [INDIRECT_PROMPT_INJECTION]: \n
Ingestion points: User-generated prose and prompts are extracted directly from local agent transcript files and databases by scripts/extract_prompts.py.\n
Boundary markers: The instructions for synthesizing the assessment and writing the report letter do not specify the use of strict delimiters or protective markers when processing the extracted transcript content.\n
Capability inventory: The skill has permissions to read local files, write assessment reports to the filesystem, and execute shell commands via subprocesses.\n
Sanitization: No content sanitization or filtering of potential malicious instructions within the extracted transcript prose is performed before processing by the agent.\n- [COMMAND_EXECUTION]: The scripts collect.py and git_stats.py utilize subprocess.run to execute local system commands, including python3 for script orchestration and git for mining contribution statistics from local repositories.

assess-me