The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes untrusted data from the screen (via OCR) and can execute UI actions based on that data.
Ingestion points: The image_snapshot and image_find_text tools in SKILL.md ingest screen content.
Boundary markers: There are no explicit prompt boundaries for the OCR results.
Capability inventory: Scripts can perform UI interactions including image_click, image_type, and image_send_keys.
Sanitization: The skill includes a safety mechanism that blocks clicks on potentially destructive UI elements (e.g., 'Delete', 'Pay') without user confirmation.
[DYNAMIC_EXECUTION]: The skill includes a tool image_write_templates that generates and writes base64 image data and constant definitions into a main.ts file as part of the standard RPA flow creation process.

exploring-image