The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to wrap the arkcli command-line interface for model management. It provides structured guidance for subcommands such as list, get, upload, update, delete, and quantize. The instructions emphasize safe execution patterns, such as using --dry-run for previews and requiring user confirmation before performing write or delete operations.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes output from external command executions and interacts with user-defined resources like TOS buckets and model metadata.
Ingestion points: CLI output from arkcli commands (JSON, CSV, and Table formats) as described in SKILL.md and reference documents.
Boundary markers: The skill does not define string-based boundary markers for its own prompts but relies on structured data formats (--format json) to separate control and data planes.
Capability inventory: The skill utilizes arkcli for cloud resource management and local tools like jq or sed for data parsing. It can perform file system reads to inspect reference documentation.
Sanitization: The skill mandates explicit validation steps, such as checking the active_endpoints field via arkcli models custommodel get --transform active_endpoints before allowing a deletion, which reduces the risk of unintended consequences from malicious or malformed model metadata.

arkcli-custommodel