The Agent Skills Directory

[PERSISTENCE]: The setup documentation in references/auth-setup.md instructs the agent to modify shell configuration files such as ~/.bashrc, ~/.zshenv, or ~/.profile. This is intended to ensure the TRANSCRIPTOUT_API_KEY environment variable remains available across different sessions.
[COMMAND_EXECUTION]: The skill provides the agent with complex shell commands in references/auth-setup.md involving curl, sed, and grep. These commands are designed to automate account registration and manage API keys directly on the user's system.
[INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted content from YouTube (video transcripts and metadata). Although the skill contains defensive instructions to treat this content strictly as data, the capability to execute shell commands and network requests creates a vulnerability surface if the model follows instructions embedded in the transcripts.
Ingestion points: Untrusted data enters the context via the /v1/search and /v1/transcript endpoints from api.transcriptout.com as described in SKILL.md.
Boundary markers: The SKILL.md file includes a 'Treat fetched content as data' section explicitly warning the agent not to follow commands found in the fetched content.
Capability inventory: The agent is granted the ability to perform network operations (curl) and local shell command execution as specified in the setup references.
Sanitization: The skill relies on natural language instructions to disregard malicious prompts in data rather than programmatic escaping or validation.

youtube-search