The Agent Skills Directory

[DATA_EXFILTRATION]: The skill aggregates and summarizes highly sensitive data such as screen activity and meeting transcripts via Luci MCP. It then transmits these summaries to an 'external connector' configured in CLAUDE.md or AGENTS.md. This architecture allows for the exfiltration of summarized private information to remote servers if the configuration is misaligned or compromised.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface. 1. Ingestion points: Processes raw activity data (screen activity, transcripts) via Luci MCP in SKILL.md. 2. Boundary markers: No explicit delimiters or instructions to ignore embedded content are provided. 3. Capability inventory: Includes filesystem writes to the reflections/daily/ directory and network transmission via an external connector. 4. Sanitization: No sanitization or validation of ingested activity data is performed.
[COMMAND_EXECUTION]: The skill utilizes shell execution (Bash) to determine the current date. While the specific usage is restricted, the capability to execute system commands is active.
[CREDENTIALS_UNSAFE]: The skill expects configuration for external connectors to be stored in project index files like CLAUDE.md or AGENTS.md. This practice encourages the storage of potentially sensitive connection strings or API endpoints in plain-text documentation files.

distill