The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation instructs users to download and install the platform CLI and various development tools. This includes the vendor's CLI from cli.inference.sh, the uv package manager from astral.sh, and Node.js version managers from fnm.vercel.app and nvm-sh. It also references numerous Python and Node.js dependencies found in standard registries.\n- [REMOTE_CODE_EXECUTION]: The skill provides installation commands that pipe remote scripts directly into a shell or execute them via PowerShell. These include curl -fsSL https://cli.inference.sh | sh, curl -LsSf https://astral.sh/uv/install.sh | sh, and curl -fsSL https://fnm.vercel.app/install | bash. These commands target the official platform domain or established development tool providers.\n- [COMMAND_EXECUTION]: The skill utilizes a custom CLI (belt) for application lifecycle management, including scaffolding (belt app init), testing (belt app test), and deployment (belt app deploy). These commands are standard for the described workflow.\n- [INDIRECT_PROMPT_INJECTION]: The skill provides templates for building applications that ingest and process user-supplied data (e.g., a prompt field in AppInput or RunInput). These applications are designed with capabilities such as network access via httpx and local file system operations. The provided templates do not include explicit sanitization or boundary markers to prevent the underlying model from potentially following instructions embedded within the processed data.\n
Ingestion points: AppInput (Python) and RunInput (Node.js) schemas in the logic templates.\n
Boundary markers: Not present in the code examples.\n
Capability inventory: Applications incorporate network request clients (httpx.AsyncClient) and file system access (reading and writing to /tmp).\n
Sanitization: Schemas provide type and structure validation via Pydantic or Zod, but no specific LLM instruction filtering is implemented.

building-inferencesh-apps