The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted "state" data (evidence) which is then judged by an AI model. This creates a potential surface for indirect prompt injection.
Ingestion points: Untrusted data enters the context via the state object in the models.classify function, as described in SKILL.md.
Boundary markers: The documentation recommends using structured JSON with named fields and backticked paths to help the model distinguish between instructions and data.
Capability inventory: The skill utilizes platform-native classification tools (models.classify, models.getModelOfType). No dangerous capabilities like file system writes, subprocess execution, or arbitrary network requests are present.
Sanitization: The skill proactively addresses security by instructing users to "Write criteria for the attack," defining how the model should handle injected instructions or spoofed fields within the state data.
[EXTERNAL_DOWNLOADS]: The skill references official documentation and resources from well-known services and the vendor's own infrastructure.
Fetches documentation and implementation patterns from docs.typesafe.ai.
References official Cloudflare Workers AI documentation and Hugging Face model repositories for configuration and self-hosting guidelines.

decision-models