The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads sensitive conversation history from ~/.claude/projects/**/*.jsonl. These logs contain the full history of user interactions with the AI, which may include proprietary code, environment variables, or other sensitive configuration data disclosed during sessions.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical session data to generate instructions and automations, creating a surface for indirect injection.\n
Ingestion points: ~/.claude/projects/ session files processed by scripts/digest.mjs.\n
Boundary markers: Absent; the content is aggregated into a Markdown digest without clear delimiters to isolate historical data from current instructions.\n
Capability inventory: Modification of CLAUDE.md, creation of new agent skills in ~/.claude/skills/, and execution of generated scripts.\n
Sanitization: None performed on the extracted historical text.\n- [DYNAMIC_EXECUTION]: The skill is designed to generate new executable scripts or automations based on its analysis and execute them on the user's machine.\n- [COMMAND_EXECUTION]: Invokes node to run a local script (scripts/digest.mjs) included with the skill package to parse session history.\n- [PERSISTENCE]: The skill allows the creation of 'scheduled tasks' as a form of automation, which could be used to establish persistent code execution on the host system.

ameliore-systeme