The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to process user-supplied narration scripts (ingestion point in SKILL.md), which is a known vector for indirect prompt injection. To address this, the skill includes explicit boundary instructions directing the agent to treat input strings (including potential injection markers like 'ignore previous instructions') strictly as data to be voiced rather than instructions to be followed. The skill provides sanitization guidance by advising the agent to flag sensitive information like PII and utilizes capabilities including tool calls and shell execution within a sandbox.
Command Execution: The skill executes a provided bash script (scripts/speech_metrics.sh) and ffmpeg via the sandbox_exec interface. These operations are part of the skill's core functionality to measure speech duration and assemble audio segments to fit specific timecode requirements.
Sensitive Data Management: The skill provides clear instructions for handling narration text, advising against transmitting personally identifiable information or credentials through the external text-to-speech provider. It also follows recommended practices by directing the agent to read necessary API keys from the environment rather than including them in prompts or logs.

narrator