The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads browser binaries and voice models during setup from remotion.media and huggingface.co. These are well-known and trusted services required for the skill's rendering and text-to-speech functionality.
[COMMAND_EXECUTION]: The skill executes local commands including npm, ffmpeg, ffprobe, and git. These are used for standard development tasks such as dependency installation, video/audio processing, and extracting project history.
[CREDENTIALS_UNSAFE]: The skill reads API keys for ElevenLabs and Fish Audio from the environment or .env files. It implements safety checks, such as verifying if .env is listed in .gitignore, and includes explicit instructions to never print, commit, or exfiltrate these keys.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from project files (e.g., DESIGN.md, README.md). It contains defensive instructions to ignore any commands or requests for secrets that might be embedded in these files.
[DATA_EXFILTRATION]: Narration text is sent to ElevenLabs or Fish Audio for speech synthesis when the user opts for those services. This is a documented, user-invoked feature, and no unauthorized exfiltration of sensitive data was detected.

design-handoff