The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructs the agent to execute yt-dlp --cookies-from-browser <browser_name>. This command harvests sensitive session cookies and authentication data from the user's local browser to bypass YouTube's anti-bot measures, creating a high risk of credential exposure.
[COMMAND_EXECUTION]: The shell command sequence provided for transcript extraction (yt-dlp ... "<URL>" ; mv raw_transcript.*.json3 raw_transcript.json) is vulnerable to command injection. If the user-provided URL contains shell metacharacters like semicolons or pipes, the agent may execute arbitrary system commands.
[COMMAND_EXECUTION]: The "Subagent Dispatch Contract" section contains obscured tool and command names using the BEL control character (\u0007) and newline characters (\n). Specifically, the text contains \u0007sk_permission and \nead_file (intended as read_file). This obfuscation is a characteristic indicator of attempts to evade security filters that monitor for privileged command usage.
[PROMPT_INJECTION]: The skill employs strong imperative language and "Iron Rules" to override agent behavior. Furthermore, it is vulnerable to indirect prompt injection via processed transcript data.
Ingestion points: External YouTube transcript data fetched via yt-dlp (SKILL.md).
Boundary markers: Absent. The skill summarizes transcript text without using delimiters or instructions to ignore embedded commands.
Capability inventory: Shell execution, file manipulation, and permission request tools (sk_permission).
Sanitization: No sanitization or filtering of the transcript text is mentioned before processing or summarization.
[DATA_EXFILTRATION]: The skill combines the ability to access local browser cookies with the ability to write files and potentially access external URLs during the "Verify reference links" step. This creates a risk where harvested session data could be exfiltrated through the skill's normal operations.

youtube-minutes-synthesis