The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/fetch_youtube_transcript.py executes the yt-dlp command via subprocess.run to extract transcripts from YouTube videos. Additionally, scripts/validate_contract.py uses shell commands to execute and test internal scripts during the setup or validation process.
[EXTERNAL_DOWNLOADS]: The skill fetches data from external platforms including YouTube and X (Twitter) to generate daily research reports. This involves downloading transcripts and web content for processing.
[PROMPT_INJECTION]: The skill processes untrusted data from external sources (tweets and transcripts), creating a surface for indirect prompt injection. However, this is mitigated by robust instructions in SKILL.md and references/research-bridge.md which mandate that all findings be classified as 'source leads' requiring manual verification. The skill also explicitly prohibits the use of auto-translated UI text or unexpanded summaries as valid evidence.
Ingestion points: scripts/fetch_x_daily.py (X/Twitter posts) and scripts/fetch_youtube_transcript.py (YouTube transcripts).
Boundary markers: The skill instructions require the agent to wrap findings in 'Source Leads' tables and include explicit 'requires-primary-check' caution labels.
Capability inventory: The agent has the capability to write to the local research dossier (companies/, modules/, etc.) via scripts/route_research_updates.py.
Sanitization: Basic text normalization is performed in scripts/creator_tracker_lib.py, and the agent is instructed to manually verify claims against primary sources before updating the core research model.

creator-research-tracker