The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill workflow relies on npx --yes hyperframes@0.8.70 to download and execute code from a third-party package not included in the trusted vendor list. This tool is used for visual verification, frame snapshotting, and final video rendering.
[EXTERNAL_DOWNLOADS]: The record.mjs script utilizes @puppeteer/browsers to download a specific version of the Chrome for Testing binary (approx. 150MB) from Google's servers. It also triggers the installation of several Node.js dependencies from the npm registry, including @strudel/web and puppeteer-core.
[DATA_EXFILTRATION]: The hyperframes tool's snapshot command is documented to upload video frames to a remote hosted captioning model by default. While the skill instructions advise using the --describe false flag to prevent this, the inherent default behavior of the external dependency represents a data exposure risk.
[INDIRECT_PROMPT_INJECTION]: The skill processes .strudel.js files, which contain executable JavaScript code. This code is evaluated in a browser environment via Puppeteer during the recording phase. If an agent composes the track based on untrusted user input without sanitization, it could lead to the execution of malicious logic within the browser sandbox.
[COMMAND_EXECUTION]: The skill requires the execution of multiple shell commands (e.g., npm install, node record.mjs, npx hyperframes) which could be misused if project paths or configuration files are manipulated to point to locations outside the intended project directory.

prod-by-claude