The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted caption and transcript data from external YouTube videos. This data is processed by the agent and stored in Markdown files, creating a surface for indirect prompt injection if the captions contain instructions intended to override the agent's behavior.
Ingestion points: Transcript fetching occurs in scripts/fetch_transcript.py via the youtube-transcript-api library.
Boundary markers: The output Markdown uses standard YAML delimiters (---) for frontmatter, but the skill lacks specific delimiters or "ignore" instructions for the agent to separate the fetched transcript content from its own instructional context during summarization.
Capability inventory: The skill has permissions to write to the local filesystem and execute the yt-dlp binary via subprocess calls.
Sanitization: The script uses JSON-based quoting for metadata fields and basic Markdown table sanitization, but the transcript text itself is not filtered or sanitized for potential injection attempts.
[COMMAND_EXECUTION]: The script scripts/fetch_transcript.py invokes the yt-dlp command-line tool using subprocess.run to retrieve video metadata. The tool invocation is considered safe as the video ID input is strictly validated against a regular expression before use.
[EXTERNAL_DOWNLOADS]: Documentation and planning files within the skill reference codeguilds.dev. Automated scanners have flagged this domain as being on a reputation blacklist, posing a potential risk if the site is used for further distribution or package hosting.
[DYNAMIC_EXECUTION]: The dependency check function in scripts/fetch_transcript.py uses importlib.import_module to verify the presence of required libraries. This is implemented safely using hardcoded string literals for module names.
[METADATA_POISONING]: Automated scanners flagged the SKILL.md file with a reputation alert (FileRepMalware). While static analysis of the file shows benign instructional content, the alert suggests the metadata or phrasing may trigger reputation-based security filters.

youtube-fetcher