The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes interactable elements of web pages to determine the agent's next actions. This creates an attack surface where a malicious web page could attempt to influence the agent's behavior through crafted element names or labels.
Ingestion points: Web page content (interactable elements) is extracted in snapshot.js and passed to the agent context in discovery.js for decision-making.
Boundary markers: The prompt in discovery.js does not use robust delimiters or specific instructions to ignore malicious content within element names or attributes.
Capability inventory: The skill can perform automated browser actions (click, type, scroll), execute shell scripts, and upload content to external services via the moda CLI.
Sanitization: The skill relies on accessibility tree data and attributes (like aria-label) extracted by Playwright.
[DYNAMIC_EXECUTION]: The script compile.py dynamically modifies the Python module search path (sys.path) to import code from a local project directory (a "studio" checkout). This behavior is intended for integration with local development environments but involves loading and executing code from computed paths.
[EXTERNAL_DOWNLOADS]: During installation, the skill's package configuration uses npx to download and install the Chromium browser for Playwright. This is a standard operation for browser automation libraries and originates from a well-known service provider (Microsoft).
[COMMAND_EXECUTION]: The skill frequently executes shell commands and Node.js scripts (e.g., run.mjs, doctor.mjs, ffmpeg, and the moda CLI) to perform its core recording, processing, and publishing tasks.

moda-video-demo