The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external code files and conversation history to generate explanations, creating a potential surface for indirect instructions. 1. Ingestion points: The agent reads local source code files, accesses thread history via t3_thread_read, and processes results from delegated how and why skills. 2. Boundary markers: The instructions lack explicit delimiting or 'ignore embedded instructions' directives for the ingested content. 3. Capability inventory: The skill utilizes delegate_task to invoke other skills and uses preview_* and device_* tools to interact with web environments or simulators. 4. Sanitization: There is no mention of sanitizing, escaping, or validating the ingested external data before it is used to generate the final explanation.

teach