The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The ShowNotesGenerator class incorporates untrusted transcript data directly into LLM prompts without using delimiters or sanitization. This creates a surface for indirect prompt injection if a podcast transcript contains instructions designed to manipulate the AI's summary or highlight generation.
Ingestion points: The transcript parameter in the _generate_description, _extract_highlights, and _extract_resources methods within SKILL.md.
Boundary markers: Absent; the transcript is interpolated directly into the f-string prompt.
Capability inventory: The LLM is used to generate text (descriptions, highlights, and resource lists) based on the input.
Sanitization: None; the transcript text is used as provided.
[COMMAND_EXECUTION]: The skill uses the subprocess module to perform audio manipulation via the ffmpeg CLI.
The PodcastEditor class in SKILL.md constructs complex filter strings for ffmpeg using subprocess.run.
While the command is executed using a list of arguments (reducing shell injection risk), the filter_complex string is built using variables derived from user-provided edit timings.
[EXTERNAL_DOWNLOADS]: The skill relies on and references several third-party libraries for core functionality.
It utilizes openai-whisper for audio transcription.
It utilizes pyannote.audio for speaker diarization, which requires an authentication token.
It utilizes feedgen for RSS XML generation and websockets for a synchronization server used in remote recording.

podcast-production