The Agent Skills Directory

Indirect Prompt Injection: The skill is designed to ingest and process external source material such as newsletters, digests, or articles to generate podcast scripts.\n
Ingestion points: Untrusted data enters the agent context through the source and topic inputs defined in SKILL.md.\n
Boundary markers: The instructions do not specify the use of clear delimiters or explicit 'ignore' commands for instructions embedded within the processed material, although it does include steps to filter boilerplate content.\n
Capability inventory: The agent has the capability to write files to the /app/created/ directory, execute Python code for audio processing, and call the ConverttexttospeechwithSSML tool.\n
Sanitization: The skill provides logic to sanitize the slug used in filenames (lowercase-hyphenated, ASCII only), but lacks explicit sanitization or escaping for the core content processed by the model.\n- Dynamic Execution: The skill provides a Python template for the agent to decode base64 audio data and instructs it to write logic for stitching audio files using the wave module. While essential for generating the podcast audio as intended, the generation and execution of scripts at runtime represents a security surface area.\n- File System Interaction: The skill writes transcript, SSML, and audio files to the local file system. The instructions include a sanitization step for the user-derived slug to mitigate potential path traversal risks during file creation.

generating-podcast-script