The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script scripts/video_to_obsidian.py contains a hardcoded API key for the YouTube InnerTube API (_IK). The key is constructed using string concatenation ("".join(...)), which is a technique used to evade simple static detection. While identified by the author as a public client key, hardcoding credentials and using string obfuscation are generally unsafe practices.
[COMMAND_EXECUTION]: In SKILL.md (Step 6), the instructions direct the AI agent to execute shell commands using os.system() to open Obsidian deep links. These commands interpolate variables like rel_path directly into a shell string. Since the safe_filename function in the supporting script only filters filesystem-illegal characters and does not sanitize shell metacharacters (such as $, &, ;, or |), a malicious YouTube video title could lead to command injection when the agent attempts to open the resulting file.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data (YouTube titles, descriptions, and transcripts) and uses it to populate Markdown templates and guide agent reasoning.
Ingestion points: YouTube metadata retrieved via urlopen and transcript content read from local files generated by the TranscribeYoutube skill.
Boundary markers: The instructions do not define clear boundary markers or isolation protocols to prevent instructions embedded in video content from influencing the agent's logic.
Capability inventory: The skill possesses file system read/write capabilities, network access, and the ability to execute subprocesses.
Sanitization: Sanitization is limited to basic filename cleaning and does not account for instructional content or shell-sensitive characters in the data stream.

VideoToObsidian