The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill processes data from external sources (database bookmarks) and interpolates it directly into the prompt for audio generation. This represents a potential surface for indirect prompt injection if the source content is attacker-controlled.
Ingestion points: The generate_audio_narrative method in references/code-examples.md retrieves bookmark titles and summaries from a database.
Boundary markers: The prompt construction uses direct string interpolation without explicit delimiters (e.g., XML tags or triple backticks) or instructions to ignore embedded commands within the content.
Capability inventory: The skill facilitates WebSocket communication with Azure OpenAI's Realtime API to generate audio and transcripts, and performs database write operations.
Sanitization: The implementation does not show explicit sanitization or validation of the bookmark content before it is processed by the AI model.
[Secure Credential Management]: The skill follows security best practices by instructing users to manage sensitive credentials like AZURE_OPENAI_AUDIO_API_KEY via environment variables rather than hardcoding them in the application logic.
[Safe Audio Data Processing]: The audio conversion utility in scripts/pcm_to_wav.py uses standard Python libraries (struct, io) to perform deterministic format conversion (PCM to WAV) without the use of dynamic execution or risky deserialization patterns.

podcast-generation