The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs users to install the platform CLI using a piped shell command: curl -fsSL https://cli.inference.sh | sh. This is a standard installation pattern for the official tool from the vendor's primary domain.
[EXTERNAL_DOWNLOADS]: Several installation scripts are referenced for development prerequisites from well-known and trusted sources:
Fetches the uv installer from Astral's official site (astral.sh).
Downloads the fnm installer from Vercel's public domain (fnm.vercel.app).
Retrieves the nvm installation script from the official GitHub repository for the project.
[COMMAND_EXECUTION]: The skill documentation uses shell commands to manage the development lifecycle, including belt app init, belt app test, and belt app deploy. These are legitimate operations for the stated purpose of building apps on the platform.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill provides documentation on how to securely manage API keys and OAuth credentials using environment variables (os.environ in Python, process.env in Node.js). This follows industry best practices for application secret management.
[INDIRECT_PROMPT_INJECTION]: The skill documents how to build applications that process external inputs (prompts, files). While this introduces an attack surface for indirect prompt injection, the documentation encourages the use of Zod and Pydantic for strict schema validation and data sanitization.
Ingestion points: Data enters the app via AppInput / RunInput models defined in inference.py or inference.js.
Boundary markers: Code templates do not explicitly show boundary markers for prompt interpolation, but focus on structured data handling.
Capability inventory: Apps have the capability to perform network requests (httpx), file operations, and utilize GPU resources as defined in inf.yml.
Sanitization: Relies on schema validation through Pydantic (Python) and Zod (Node.js).

building-inferencesh-apps