The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell scripts and Python subprocess calls to orchestrate media tools (ffmpeg, ffprobe), repository management (git, gh), and development environments (npm, npx, playwright). These tools are used to trim video silence, cut spans, and verify build integrity.
[INDIRECT_PROMPT_INJECTION]: The pipeline ingests external data from Fathom transcripts and Zoom recordings to generate structured summaries and update public YouTube metadata. This creates an attack surface where malicious content in a transcript could influence downstream actions.
Ingestion points: Reads and parses FATHOM_FILE, Fathom JSON, and Zoom VTT transcript files.
Boundary markers: Missing in most scripts; the skill relies on LLM context instructions in SKILL.md to identify boundaries.
Capability inventory: Includes shell command execution (ffmpeg, git), file system writes (MDX documentation), and network API updates (YouTube Data API).
Sanitization: Implements mdx_escape in update_meeting_doc.py and rebuild_aggregations.py to filter MDX/JSX hazards such as HTML comments, brackets, and braces.
[PERSISTENCE]: The skill includes scripts/poll-and-publish.sh, which is designed to be integrated with launchd for daily automation. It utilizes marker files located in ~/.claude/state to track execution state across sessions.
[CREDENTIALS_UNSAFE]: The skill requires access to sensitive local credential files, specifically ~/.zoom_credentials/oauth_token.json and token.pickle (YouTube). Instructions in SKILL.md and references/youtube-api.md include Python and Bash snippets to programmatically extract and use these access tokens for authenticated API requests.
[EXTERNAL_DOWNLOADS]: The pipeline downloads video recordings from Zoom and Fathom using URLs retrieved from their respective APIs. It also uses npx playwright install chromium to fetch the browser binaries required for thumbnail rendering.

agency-docs-updater