The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external audio data from URLs or local files to generate transcripts. If this transcript content is subsequently processed by an AI agent without sanitization, it could act as a vector for indirect prompt injection.
Ingestion points: The transcribeUrl and transcribeFile methods in SKILL.md (Step 1) and references/implementation.md ingest external audio content.
Boundary markers: No delimiters or "ignore embedded instructions" warnings are included for the generated transcript text.
Capability inventory: The skill facilitates reading external data and returning the resulting text to the agent's context.
Sanitization: No transcript validation or sanitization logic is provided in the implementation templates.
[EXTERNAL_DOWNLOADS]: The skill documentation references the installation of standard third-party libraries for audio transcription and concurrency management.
Evidence: Mentions @deepgram/sdk and p-limit for Node.js environments, and the deepgram-sdk for Python environments. These are well-known packages for the intended functionality.

deepgram-core-workflow-a