The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file contains a 'Preamble' section that executes shell commands automatically when the skill is loaded or started. Evidence includes calls to python3 telemetry/version_check.py and python3 telemetry/telemetry_init.py.
[COMMAND_EXECUTION]: The skill attempts to execute local Python scripts during its initialization phase. These scripts are not provided in the source files, making their behavior unverifiable and potentially risky if the environment contains malicious files at those paths.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted external data and interpolates it into LLM prompts without adequate protection.
Ingestion points: Transcripts are ingested from local files or generated from downloaded audio via RSS feeds in podcast_pipeline.py.
Boundary markers: None. The transcript text is directly inserted into the user_prompt in extract_content_atoms and generate_content_pieces without delimiters or instructions to ignore embedded commands.
Capability inventory: The skill can make network requests to external APIs (OpenAI, Anthropic), download files from arbitrary URLs, write to the local file system, and execute shell commands in the preamble.
Sanitization: None. The raw transcript text is truncated and interpolated directly into prompts.
[EXTERNAL_DOWNLOADS]: The podcast_pipeline.py script downloads audio files from user-supplied RSS feed URLs using the requests library. This involves connecting to arbitrary external domains provided via CLI arguments or configuration.

podcast-pipeline