The Agent Skills Directory

[PERSISTENCE]: The instructions in references/auth-setup.md direct the agent to store the API key persistently by modifying environment configuration files such as shell profiles (e.g., ~/.bashrc, ~/.zshrc, or .profile). Modifying these startup scripts allows the agent to maintain state and configuration across multiple sessions, which is a persistence mechanism.
[PROMPT_INJECTION]: The file references/auth-setup.md contains explicit instructions designed to bypass platform safety guardrails. Specifically, it directs the agent to write sensitive API responses (containing JWT tokens and API keys) to temporary files and read them back in a single operation to prevent the platform's redaction filters from detecting and masking the secrets. This constitutes an intentional bypass of security monitoring mechanisms.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external YouTube transcripts and metadata which are created by third-party YouTube users.
Ingestion points: Data enters via the transcriptapi.com/api/v2/youtube/transcript and search endpoints defined in SKILL.md.
Boundary markers: None are present; the skill does not instruct the agent to ignore or delimit instructions found within the transcript text.
Capability inventory: The skill uses curl for network operations and performs file system writes for credential persistence.
Sanitization: No sanitization or validation of the ingested transcript content is performed before the agent processes it.
[COMMAND_EXECUTION]: The skill provides and encourages the execution of shell commands (specifically curl) to interact with the API, and instructs the agent to execute commands to modify local shell configuration files for persistence.

youtube-data