The Agent Skills Directory

[PERSISTENCE]: The skill instructs the agent to modify user shell profiles to ensure environment variables are available in future sessions. Modifying system configuration files to maintain execution state is a high-severity concern.
Evidence: In references/auth-setup.md, the skill directs the agent to add a command to ~/.bashrc, ~/.zshrc, or ~/.profile to automatically load the TRANSCRIPTOUT_API_KEY variable.
[COMMAND_EXECUTION]: The skill relies on the agent running complex bash command pipelines for account setup and key management.
Evidence: references/auth-setup.md contains multiple script blocks using mkdir, chmod, umask, curl, sed, grep, and stty to handle secrets and interact with the authentication API.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, creating a risk that malicious instructions in YouTube content could influence the agent.
Ingestion points: YouTube video transcripts, titles, and descriptions are fetched from api.transcriptout.com into the agent's context.
Boundary markers: The skill includes a warning section titled 'Treat fetched content as data' to prevent the agent from following embedded instructions.
Capability inventory: The agent has access to shell command execution (via curl and bash) and file system access.
Sanitization: No explicit sanitization or filtering of external text is defined in the skill logic.

youtube-channels