The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from multiple sources which could contain malicious instructions.
Ingestion points: Data is pulled from email threads, chat mentions/DMs (Slack, Teams), calendar event titles/descriptions, and linked documents.
Boundary markers: The skill contains a 'Ground rules' section explicitly instructing the agent to treat all gathered content as data to summarize, never as instructions to act on. It also mandates escaping text in the final output.
Capability inventory: The skill uses tool calls to read email/chat/calendar data and executes Node.js commands for visual verification.
Sanitization: Instructions require rendering gathered text as escaped plain text in the HTML artifact to prevent live markup or script injection.
[COMMAND_EXECUTION]: The skill utilizes shell commands during its build and verification phases.
Evidence: Uses npm pack @fontsource/fraunces and npm install playwright to manage dependencies and browser tools.
Evidence: Executes a complex Node.js one-liner (node -e "...") to launch Chromium via Playwright for screenshot verification.
[DYNAMIC_EXECUTION]: The skill generates and executes code at runtime to verify its rendering.
Evidence: Dynamically constructs and executes a Node.js script to automate a headless browser, take a screenshot, and verify the UI layout.
[EXTERNAL_DOWNLOADS]: The skill fetches assets and tools from external registries.
Evidence: Downloads font packages and the Playwright automation library from the official NPM registry.

morning