The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill includes instructions for installing several CLI tools (belt, uv, fnm, nvm) by downloading and executing scripts via curl | sh, curl | bash, or PowerShell's Invoke-Expression. These downloads originate from trusted organizations (Vercel), well-known developer tool providers (Astral, NVM), or the vendor's own official domain (inference.sh).
[PRIVILEGE_ESCALATION]: Installation guidance for the uv tool on Windows includes a command to bypass the PowerShell execution policy (-ExecutionPolicy ByPass). This is a common requirement for installing developer tooling on Windows and is scoped to the specific installation command.
[PROMPT_INJECTION]: The instructions contain directives for the AI agent to prioritize the platform's CLI-based scaffolding over local documentation or README files. This pattern is used to ensure the agent follows the correct deployment workflow and avoids common manual configuration errors.
[INDIRECT_PROMPT_INJECTION]: The skill provides templates for creating applications that process external user data (prompts and files). While this creates a potential surface for indirect injection in the resulting apps, the skill itself provides standard development patterns for handling such inputs.
Ingestion points: AppInput (Python) and RunInput (Node.js) schemas defined in inference.py and src/inference.js.
Boundary markers: Not explicitly defined in the templates; the skill assumes standard validation via Pydantic and Zod.
Capability inventory: The documented applications have capabilities including file system access (writing to /tmp), network requests (via httpx or OpenAI clients), and access to secrets stored in environment variables.
Sanitization: The provided examples rely on platform-level validation (Zod/Pydantic) but do not include specific natural language sanitization logic.

building-inferencesh-apps