The Agent Skills Directory

[DATA_EXFILTRATION]: The skill uses ElevenLabs Scribe v2 for cloud-based transcription. This involves transmitting user media to an external service. However, the skill implements a robust consent mechanism, explicitly requiring the agent to notify the user about the destination, potential costs, and quota usage before seeking explicit approval for each upload.
[COMMAND_EXECUTION]: The skill relies on system binaries including ffmpeg and ffprobe for media analysis and rendering. Safety is maintained by strictly prohibiting silent installation or modification of these tools and ensuring that the source media is never overwritten, moved, or deleted during the process.
[DATA_EXPOSURE]: All generated artifacts and work files are confined to a specific edit/ subdirectory adjacent to the source media. This strict output contract prevents unauthorized access to other directories and ensures clear isolation between the original project files and the skill's outputs.
[PROMPT_INJECTION]: As a tool that processes external media, the skill presents an indirect prompt injection surface where audio or text within the video could contain malicious instructions. The skill mitigates this through a mandatory evidence chain:
- Ingestion points: User-provided video/audio files processed in SKILL.md (Step 1).
- Boundary markers: Mandatory user consent for cloud uploads and a required 4–8 sentence strategy approval step before any editing occurs.
- Capability inventory: Media processing via ffmpeg/ffprobe, image generation via Pillow, and cloud transcription via ElevenLabs API.
- Sanitization: Use of cached word-level verbatim timestamps to ensure transcription accuracy and a mandatory three-pass QA process to verify rendered outputs before delivery.

chatgpt-short-video-editor