The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the user's screen through the screenshot tool call.
Ingestion points: The skill instructs the agent to capture screenshots using curl to interact with a local API in SKILL.md.
Boundary markers: Absent. There are no instructions to the agent to distinguish between UI elements and potentially malicious text content displayed within the captured windows.
Capability inventory: The skill can execute shell commands via curl to interact with the local bridge, including showing overlays (openclicky_point), speaking text (speak), and capturing more screenshots.
Sanitization: Absent. No mention of filtering or sanitizing text extracted or interpreted from images.
[COMMAND_EXECUTION]: The skill relies on curl commands to interact with a local bridge service for task execution.
Evidence: Several examples in SKILL.md demonstrate using curl to perform health checks, discover tools, and call specific bridge functions via JSON POST requests to http://127.0.0.1:32123.

openclicky-guided-tutorials