The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download the runpodctl binary from the official RunPod GitHub repository (github.com/runpod/runpodctl). This is a well-known and trusted source for this infrastructure-related tool.
[COMMAND_EXECUTION]: The deployment workflows involve executing shell commands for tool installation and project deployment, including the use of sudo to move the CLI binary into the system's executable path (/usr/local/bin/). This represents standard developer-level environment setup.
[INDIRECT_PROMPT_INJECTION]: The provided Python worker templates ingest untrusted data from the job["input"] object, which is then passed to an LLM.
Ingestion points: External input enters the system through the prompt and messages fields in templates/runpod-worker.py and various code snippets in SKILL.md.
Boundary markers: No explicit delimiters or boundary markers are used in the provided template code to separate user input from system instructions.
Capability inventory: The worker handlers possess capabilities for network communication (via the runpod SDK and requests) and filesystem operations (for logging and model checkpointing).
Sanitization: Input validation is performed via pydantic for schema correctness, but no natural language sanitization is applied to the content of the prompts themselves.
[DYNAMIC_EXECUTION]: The skill references the use of torch.load in reference/pod-management.md for managing model checkpoints. While torch.load uses unpickling which can execute arbitrary code, it is a standard practice in machine learning for loading user-generated state dictionaries.

runpod-deployment