The Agent Skills Directory

[SAFE]: The skill instructions and reference materials follow security best practices by defining strict boundaries for agent behavior and including explicit 'Stop Conditions' to prevent hallucinations or unsafe execution.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an entry point for user requests, creating a surface where untrusted data could attempt to influence the routing logic. However, the skill includes significant mitigations such as the 'Routing Law' and 'Stop Conditions' which require the agent to validate and re-state the user's request against a predefined domain map before proceeding.
Ingestion points: The skill ingests user requests via the default prompt in agents/openai.yaml and the workflow in SKILL.md.
Boundary markers: Present. The skill mandates restating the requested outcome and mapping it to specific surfaces defined in references/routing-map.md.
Capability inventory: Low risk. The skill is limited to text generation and routing; it does not possess capabilities for network operations or arbitrary command execution.
Sanitization: The skill enforces strict matching against a domain ownership table, which acts as a logical filter against incidental or malicious vocabulary in the request.

using-forge