The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided text to perform speech synthesis and Hebrew vocalization (niqqud). This provides a potential surface for indirect prompt injection where malicious instructions in the script could influence the agent's behavior during the preparation phase.
Ingestion points: User-supplied narration, dialogue turns, and script text processed in SKILL.md and references/voices/speech-localization.md.
Capability inventory: Tools for generating speech, designing voices, and merging media across providers like ElevenLabs and Cartesia.
Boundary markers: The skill instructs the agent to preserve wording and meaning, but does not provide specific technical delimiters to separate user text from agent instructions.
Sanitization: No explicit mention of input sanitization or filtering for malicious instructions in the text payload before transmission to external tools.
[EXTERNAL_DOWNLOADS]: The skill retrieves model parameters and voice catalogs from several well-known external AI services, including ElevenLabs, Cartesia, Fish Audio, HeyGen, xAI, and MiniMax. These references are consistent with the skill's primary functionality and target established providers.
[DATA_EXFILTRATION]: The skill implements an optional skills_used field for workflow analytics, which transmits usage information (skill and guide names) to the vendor's platform. The instructions explicitly prohibit the inclusion of sensitive data, such as prompts, local file paths, or conversation history, within this field.

creativeclaw-generate-voiceover