The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements the Sprocket worker framework which is designed to process untrusted data payloads from asynchronous task queues.
Ingestion points: The predict(self, args: dict) method in scripts/sprocket_hello_world.py and references/sprocket-sdk.md processes a payload directly from the queue client.
Boundary markers: No explicit delimiters or boundary markers are enforced by the framework to isolate instructions from data within the JSON payload.
Capability inventory: The skill enables broad capabilities including container build-time shell command execution (run instructions in pyproject.toml), runtime file operations including automated uploads (sprocket.FileOutput), and network access to Together AI's API.
Sanitization: No input sanitization is provided in the library or templates; developers are responsible for validating and escaping incoming job payloads before use.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of vendor-specific SDKs and well-known machine learning libraries.
Installs the sprocket SDK from the vendor's private registry at https://pypi.together.ai/.
Recommends fetching CUDA-enabled PyTorch wheels from the official pytorch.org index to support GPU workloads.
[COMMAND_EXECUTION]: The Jig CLI configuration in references/jig-cli.md allows users to define arbitrary shell commands within the pyproject.toml file under [tool.jig.image] run and cmd keys. These commands are executed by the together CLI tool during the Docker image build and container startup phases.

together-dedicated-containers