The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the holoscan-cli tool and project templates from official package registries including PyPI and the NVIDIA package index.
[EXTERNAL_DOWNLOADS]: Clones project source code and configuration from the official NVIDIA Holoscan GitHub organization.
[COMMAND_EXECUTION]: Uses standard development utilities including uv, git, cmake, and dpkg to perform module scaffolding, building, and packaging tasks as intended by the skill's primary purpose.
[INDIRECT_PROMPT_INJECTION]: The skill processes external module metadata, API contracts, and source code which could serve as an attack surface for indirect injection.
Ingestion points: Reads external configuration files including metadata.json, pyproject.toml, and module source code in SKILL.md and referenced development guides.
Boundary markers: While not using explicit LLM-specific boundary markers, the instructions emphasize isolating task-owned prefixes and working directories.
Capability inventory: Includes shell command execution via the holoscan CLI and standard build tools, file system writes during scaffolding, and network access for cloning repositories.
Sanitization: The instructions mandate schema validation of metadata against the official Holoscan schema and manual inspection of generated code and Git status before execution or committing.

holoscan-module-lifecycle