The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from platform APIs, web pages, and account exports. It mitigates this risk by explicitly instructing the agent to classify this content as untrusted and to ignore any instructions embedded within it. Ingestion points include API responses, web screenshots, account exports, and competitor pages. Boundary markers are established via explicit instructions in SKILL.md. Capability inventory includes writing logs and reports to local directories and executing defined local scripts. Sanitization involves validating input completeness and using bounded workers.
[COMMAND_EXECUTION]: The skill facilitates command execution through localized scripts (e.g., scripts/generate_image.py) and platform-specific CLI tools. Security is maintained through a 'mutation gate' that requires human approval, idempotency keys, and explicit object IDs before any state-changing operation is performed.
[DATA_EXFILTRATION]: While the skill interacts with multiple external advertising platforms, it contains strict policies against storing or logging credentials, cookies, or tokens. It mandates the use of environment variables or OS keychains for secret management.

ads