The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a pipeline that ingests external data in stage S0, including video links and user-provided requirements. This data is subsequently processed by multiple sub-skills and local Python scripts in stages S1 through S5. Malicious content embedded in the source media or metadata could potentially attempt to influence the agent's behavior or the execution of downstream tools.
Ingestion points: Video links and user requirements ingested in SKILL.md (Stage S0).
Boundary markers: The workflow includes multiple manual validation gates (G2, G3, G6) where a user must verify intermediate results (DNA cards, lyrics, and audio alignment) before the process continues.
Capability inventory: The skill orchestrates file system writes to specific project directories and executes several local Python scripts (scripts/measure_tempo.py, scripts/tts.py, etc.) alongside audio processing tools like demucs and ffmpeg.
Sanitization: The high-level orchestrator does not explicitly define sanitization logic, relying on the sub-skills and human review stages to catch anomalies.
[COMMAND_EXECUTION]: The workflow relies on the execution of several local Python scripts located in the scripts/ directory and system-level audio processing utilities (demucs, faster-whisper, RubberBand). These operations are well-defined within the context of the skill's primary purpose as a media processing pipeline.
[CREDENTIALS_UNSAFE]: The skill references the environment variable DOUBAO_SPEECH_API_KEY for interacting with the Volcengine/Doubao TTS API. Accessing credentials through environment variables is a secure alternative to hardcoding secrets, which protects the integrity of the user's API keys.

rap-clone-workflow