The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit instructions in references/brief-template.md on how to bypass safety filters of downstream video generation models. It suggests substituting prohibited emotional descriptors (e.g., 'tense', 'suspenseful') with visual approximations to circumvent content controls.
[COMMAND_EXECUTION]: The scripts/editorial_pass.sh script executes shell commands and ffmpeg with complex filter chains to perform multi-stage video production tasks, including clipping, color grading, and audio mixing.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to download audio assets from Pixabay's content delivery network and communicates with ElevenLabs' official API for voice synthesis. These are documented as part of the primary workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied narrative concepts to generate prompts for external AI services. This pipeline lacks robust input validation or boundary markers between user-controlled data and instruction templates, creating a surface for indirect prompt injection.
Ingestion points: User mentions of trailer concepts, taglines, and scene descriptions in SKILL.md.
Boundary markers: None identified.
Capability inventory: Shell command execution (ffmpeg), network access (ElevenLabs, Pixabay), and local file system writes.
Sanitization: No escaping or validation of user-provided narrative content before interpolation into downstream prompts.
[CREDENTIALS_UNSAFE]: scripts/synthesize_vos.py and SKILL.md reference a specific local directory structure (~/Documents/Skills LLMS/Claude/Skills/.heygen-credentials/) for storing API keys. While the keys themselves are not hardcoded, the disclosure of this specific path and the use of a simple text file for secret storage is a metadata exposure concern.

cinematic-trailer-pipeline