The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from conversation replies and user messages to generate highlights and summaries.
Ingestion points: The readReply function in hooks/register.tsx ingests text from the assistant's last reply and a sample of the user's previous message.
Boundary markers: User content is wrapped in triple angle brackets (<<<...>>>), and assistant replies are formatted as a numbered list of paragraphs to provide structure for the summarizing model.
Capability inventory: The results of the model analysis are limited to modifying the background color of paragraphs in the UI (backgroundColor) and displaying a toast notification ($.ui.toast). The skill does not have capabilities for file writing, network requests to external domains, or command execution based on this data.
Sanitization: The skill employs a parseAnswer function that strictly extracts and validates JSON from the model's response. It also uses keepMarks to ensure highlighted indices are within valid ranges and do not target sensitive elements like code blocks or headings.

tldr-highlight