The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, specifically transcripts and video frames fetched via tools like watch-video or social-fetch. The instructions in references/edit-anatomy.md explicitly recognize this as a prompt-injection surface and direct the agent to analyze the content without following embedded instructions.
Ingestion points: references/edit-anatomy.md (via watch-video and social-fetch).
Boundary markers: The skill advises a "review once" gate and explicit warnings to treat fetched content as untrusted.
Capability inventory: Use of Playwright for rendering, ffmpeg for video encoding, and network requests to video APIs.
Sanitization: Recommends manual review of beat sheets before execution.
[DYNAMIC_EXECUTION]: The skill utilizes browser automation (Playwright) to render HTML/CSS/JS DOM snapshots into video frames. This is a core part of its programmatic video functionality and is used within a documented, legitimate context.
[COMMAND_EXECUTION]: The skill provides instructions for using CLI tools like ffmpeg, npx hyperframes, and npx create-video. These are standard industry tools for video processing and web development, used here for their intended purposes.
[CREDENTIALS_UNSAFE]: While an example API key (sk_live_123abc) is present in the evals/evals.json file, it is used for a safety test case. The skill's instructions in references/product-demo-recording.md correctly advise users to use environment variables and to rotate any keys accidentally shared in chat.

video