The Agent Skills Directory

[SAFE]: The skill provides comprehensive and secure documentation for using the Modal serverless platform. All code examples represent standard platform operations and follow security best practices for handling API tokens and remote dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill describes building web endpoints and processing external data, which introduces a potential surface for indirect prompt injection.
Ingestion points: Web endpoints and WebSocket listeners defined in SKILL.md and references/advanced-usage.md.
Boundary markers: Code snippets show basic FastAPI structures; explicit prompt boundary markers are not demonstrated.
Capability inventory: The skill facilitates command execution via subprocess.run and provides interactive sandbox environments via modal.Sandbox.
Sanitization: Examples use standard Python and FastAPI type hinting for input validation.
[EXTERNAL_DOWNLOADS]: The skill documents the installation of necessary ML libraries and platform tools from public registries and official repositories.
Installs the modal CLI and dependencies such as torch, transformers, and accelerate.
References the official Hugging Face transformers repository for specific Git-based installations.
[COMMAND_EXECUTION]: Instructions include the use of system commands for verifying environment configurations and launching distributed training jobs.
Uses subprocess.run to call nvidia-smi for GPU health checks.
Uses torch.distributed.launch for parallel training orchestration.
[DYNAMIC_EXECUTION]: The skill provides guidance on using isolated execution environments for running arbitrary code.
Demonstrates the use of modal.Sandbox to execute Python code in a containerized environment.

modal-serverless-gpu