The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external binaries to process media files.
Evidence: Scripts spawn ffmpeg, ffprobe, and yt-dlp using node:child_process.spawnSync in scripts/figureout.mjs, scripts/frames.mjs, and scripts/whisper.mjs.
Analysis: These are well-known, industry-standard tools for media processing. The skill uses a resolveBinary utility to safely locate executables on different platforms, reducing common shell-related command injection risks.
[CREDENTIALS_UNSAFE]: The skill manages API keys for external transcription services.
Evidence: scripts/config.mjs reads GROQ_API_KEY and OPENAI_API_KEY from a local .env file.
Analysis: The skill implements standard secret management by reading keys from a user-configured .env file in the home directory (~/.config/figureout/.env) or local working directory. This is consistent with safe development practices for handling local secrets.
[DATA_EXFILTRATION]: The skill optionally sends audio data to external transcription providers.
Evidence: scripts/whisper.mjs sends extracted audio to api.groq.com or api.openai.com via global fetch.
Analysis: This behavior is the primary purpose of the transcription fallback and is clearly documented in SKILL.md and references/transcription.md. The skill only uploads audio tracks (never the full video) to these well-known services when local captions are unavailable and the user has explicitly configured an API key.
[DYNAMIC_CONTEXT_INJECTION]: SKILL.md uses dynamic injection to provide environmental context to the agent at load time.
Evidence: The script !node "${CLAUDE_SKILL_DIR}/scripts/project-context.mjs"`` is invoked at the top of SKILL.md.
Analysis: The bundled project-context.mjs script emits a JSON snapshot of the system state (binary presence, platform info, and config status) to help the agent determine if the environment is ready. This is a legitimate and documented use of dynamic context for developer tools.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (video captions and metadata) from remote URLs.
Ingestion points: Video captions and metadata are fetched from remote URLs via yt-dlp in scripts/figureout.mjs.
Boundary markers: The resulting transcript is presented to the agent within a fenced code block in the markdown report.
Capability inventory: The skill utilizes Bash, Read, and AskUserQuestion tools, and can execute shell commands via bundled scripts.
Sanitization: scripts/transcribe.mjs parses and normalizes VTT text, stripping HTML-like tags from the transcript before presentation.

figureout