The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple local Python scripts as subprocesses to perform its core functions.
Calls scripts/fetch.py to retrieve video metadata and transcripts.
Calls scripts/segments.py to slice transcripts into sections.
Calls scripts/cache.py to handle canonical JSON hashing and cache validation.
Includes a fallback mechanism that executes a Python one-liner (python3 -c) using shell interpolation for transcript filtering.
[EXTERNAL_DOWNLOADS]: The skill fetches external content from YouTube, a well-known service, to obtain video transcripts and metadata. This operation is central to the skill's stated purpose and uses local caching to minimize network activity.
[PROMPT_INJECTION]: The skill processes untrusted transcript data fetched from YouTube through multiple LLM passes (Structure extraction, Claim inventory, and Synthesis).
Ingestion points: Transcript data is stored in ~/youtube-reports/.cache/{video_id}.json and subsequently read into the agent's context during processing.
Boundary markers: The SKILL.md does not specify the use of delimiters or 'ignore' instructions for the transcript content, though it enforces strict output formatting (JSON and specific Markdown structures).
Capability inventory: The agent has the capability to execute local scripts and write files to the ~/youtube-reports/ directory.
Sanitization: There is no mention of sanitization or filtering for the fetched transcript text before it is processed by the model.
[DATA_EXFILTRATION]: Network operations are restricted to fetching video data. While the skill writes reports and caches to the home directory (~/youtube-reports/), there is no evidence of unauthorized transmission of local files or credentials.

youtube-verdict