The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from YouTube transcripts via the /v1/transcript endpoint. Ingestion points: Video transcripts, titles, and descriptions fetched from api.transcriptout.com (SKILL.md). Boundary markers: The skill contains an explicit 'Treat fetched content as data' section with instructions to ignore embedded commands (SKILL.md). Capability inventory: Subprocess execution via shell commands (curl, sed, grep), file system writes/reads, and network operations (references/auth-setup.md, SKILL.md). Sanitization: No technical sanitization is applied to the transcript content; it relies on instructional guidance for the agent.
[PERSISTENCE]: The authentication guide (references/auth-setup.md) instructs the agent or user to add an export command to shell profile files (~/.bashrc, ~/.profile, or ~/.zshenv). This is used to persist the API key across terminal sessions.
[COMMAND_EXECUTION]: The skill implements its own authentication and key management workflow using shell commands. It uses curl to interact with the registration API, and sed/grep to parse JSON responses and manage local configuration files (references/auth-setup.md).
[DATA_EXFILTRATION]: The skill reads the TRANSCRIPTOUT_API_KEY from a local configuration file (~/.config/transcriptout/key) and sends it to api.transcriptout.com. This is the intended behavior for the service, and the skill follows best practices by ensuring the key is not printed or sent to other domains.

transcript