The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data via search result snippets and the file content endpoint. The instructions demonstrate high security awareness by explicitly warning the agent that this data may contain malicious instructions (e.g., 'SYSTEM: ignore previous instructions'). It provides mitigation guidance by directing the agent to treat such content as quoted data and never execute instructions found within it.
Ingestion points: matches[].snippets in GET /v1/search and the content field in GET /v1/files/{fileId}/content.
Boundary markers: The agent is instructed to treat data as 'quoted content' and use it only for citation or summarization.
Capability inventory: File management (create/read/update/delete), search, and local utility command execution (curl, python3).
Sanitization: Risk is mitigated via explicit instructional boundaries provided to the agent.
[DYNAMIC_EXECUTION]: Bash code examples provided in the skill utilize python3 -c for processing JSON API responses and calculating CRC32 checksums using standard Python libraries (json, base64, zlib). These are routine data processing tasks and do not involve the execution of untrusted code or remote payloads.

file-management