The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents a "Form-Agent Hybrid" pattern in references/patterns.md where an agent can populate form fields via tool calls. While this is an intended design pattern, it creates a surface for indirect prompt injection where untrusted model output is ingested into application state.
Ingestion points: Form fields populated by the agent as described in the "Form-Agent Hybrid" pattern in references/patterns.md.
Boundary markers: The skill suggests visually marking agent-populated fields (e.g., with a blue border) to differentiate them from user input.
Capability inventory: The skill is documentation-focused and does not include executable scripts, network operations, or privileged file access.
Sanitization: The guidelines strongly recommend using Zod schemas for strict validation and React Hook Form for performance, which serve as primary mitigations against malformed or malicious data injection.

component-patterns