The Agent Skills Directory

[SAFE]: The skill instructions emphasize local-first processing and limit data exposure by suggesting only the minimum necessary data be sent to external tools.
[SAFE]: The workflow includes built-in safeguards, such as requiring explicit user approval for calls involving costs or external processing via sandbase_run.
[INDIRECT_PROMPT_INJECTION]: The skill processes JSON Canvas files and user inputs, creating a theoretical attack surface. However, the risk is mitigated by explicit requirements for schema validation and human-in-the-loop confirmation for external actions. Mandatory Evidence Chain: 1. Ingestion points: JSON Canvas inputs and files (SKILL.md); 2. Boundary markers: Schema validation and design system preservation rules (SKILL.md); 3. Capability inventory: sandbase_run tool for task execution (SKILL.md); 4. Sanitization: Mandatory validation of serialized files after editing (SKILL.md).

json-canvas