The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a proxy for passing user-provided or external data to third-party LLMs, which is a common surface for indirect prompt injection.
Ingestion points: The --input JSON payload in SKILL.md (e.g., the prompt and system fields) is the primary entry point for untrusted data.
Boundary markers: The skill uses structured JSON to encapsulate data, providing a logical boundary between the execution command and the data payload.
Capability inventory: The skill is authorized to use the Bash tool to execute the belt CLI command.
Sanitization: No explicit sanitization of the prompt content is documented within the skill; it relies on the safety filters of the destination models (Claude, Gemini, etc.) and the inference.sh platform.
[EXTERNAL_DOWNLOADS]: The skill references and encourages the installation of external components from the inference.sh platform.
It points to installation instructions hosted at raw.githubusercontent.com/inference-sh/skills/refs/heads/main/cli-install.md.
It recommends adding the belt-sh/cli and inference-sh/skills skills to the agent environment.

llm-models