The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to install the belt CLI using a shell-piped download from https://cli.inference.sh. As this is the official domain for the platform, the remote execution is considered a vendor-provided installation method.
[EXTERNAL_DOWNLOADS]: The documentation references installation commands for several well-known development tools:
uv installation via shell script from astral.sh.
fnm installation via shell script from vercel.app.
nvm installation via shell script from GitHub.
[COMMAND_EXECUTION]: The skill instructs the agent to execute various shell commands for project scaffolding, local testing, and deployment (e.g., belt app init, belt app test, belt app deploy). These are standard operational commands for the platform's CLI.
[INDIRECT_PROMPT_INJECTION]: The skill documents the creation of applications that process external data, such as LLM prompts and user-provided files. This establishes an attack surface where untrusted input could contain embedded instructions. The documentation provides standard patterns for handling these inputs but does not include specific sanitization logic beyond type validation with Pydantic or Zod.
Ingestion points: Processed data enters via AppInput (Python) or RunInput (Node.js) models defined in inference.py and inference.js.
Boundary markers: None explicitly required in the provided templates.
Capability inventory: Apps can perform file writes to /tmp, make network requests via httpx, and utilize GPU resources.
Sanitization: Relies on schema validation (Pydantic/Zod) for data structure integrity.

building-inferencesh-apps