The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of serverless endpoints that process untrusted user input as part of their primary function (e.g., LLM prompts, data for preprocessing).
Ingestion points: Inputs are received as keyword arguments in @app.queue functions and as JSON bodies in @app.api methods (e.g., examples/api_model_service.py, examples/queue_gpu.py).
Boundary markers: None explicitly provided in the instruction templates; the skill relies on the underlying LLM's safety features.
Capability inventory: Workers have access to network volumes for persistent storage, can perform network operations to download model weights from HuggingFace, and execute Python code in a containerized environment.
Sanitization: Not explicitly documented in the provided examples; standard practice for these services assumes the application logic or model handles content safety.
[DYNAMIC_EXECUTION]: The development workflow utilizes serialization for remote execution.
Evidence: SKILL.md (Gotcha 11) notes that rp flash dev sessions and tasks ship code pickled by the local interpreter to remote workers. While this is a standard feature of the vendor's development loop for syncing local code to remote GPUs, it involves the transmission and execution of serialized objects.

flash