The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the installation of official software from the vendor's public package registries to enable cluster management capabilities.
Evidence: Instructions in SKILL.md and references/cli.md specify installing together>=2.0.0 via uv pip and the together[cli] utility.
[COMMAND_EXECUTION]: The skill uses shell commands through the vendor CLI, kubectl, and ssh to manage the full lifecycle of GPU clusters, including provisioning, scaling, and operational troubleshooting.
Evidence: Found throughout the documentation in references/cli.md and references/cluster-management.md for operations like creating clusters, downloading credentials, and node access.
[CREDENTIALS_UNSAFE]: The skill manages sensitive authentication data, such as API keys and cluster-level configuration files (kubeconfig). It adheres to standard security practices by using environment variables and default configuration paths.
Evidence: The TOGETHER_API_KEY is retrieved from environment variables, and the get-credentials command writes sensitive authentication data to ~/.kube/config.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection because it ingests and applies user-provided infrastructure configurations and Slurm lifecycle scripts.
Ingestion points: User-provided cluster names, regions, GPU types, and custom Slurm startup scripts (init, prolog, and epilog scripts) defined in SKILL.md and references/cluster-management.md.
Boundary markers: Absent. No explicit delimiters or instructions are used to prevent the agent from being influenced by malicious content embedded within user configuration strings.
Capability inventory: High-impact capabilities including cluster creation, deletion, scaling, credential harvesting, and execution of persistent scripts on remote nodes.
Sanitization: Relies on the vendor's official SDK and CLI for input validation and command construction.

together-gpu-clusters