bs-detectorlisted

# BS Detector Fact-check a short-form video claim by claim. Not a vibe check, not a summary — every checkable assertion gets pulled out, verified, and given a verdict with a source. Four phases: **transcribe → extract → verify → score.** The work is in phase 3. Phases 1 and 2 exist so nothing slips past unchecked, and phase 4 exists so the user gets an answer instead of a wall of caveats. The failure mode to avoid is a fact-check that reads as authoritative but is really just your priors with confident formatting. If you didn't check it, say you didn't check it. ## Phase 1 — Get the transcript **Default: the bundled script.** Runs locally, no account, no API key: ```bash python <skill-dir>/scripts/transcribe.py "<url-or-path>" --out ./factchecks/<slug> ``` `<skill-dir>` is the directory holding this SKILL.md — resolve it from this file's own path (typically `~/.claude/skills/bs-detector`), **not** from the working directory. The user will usually be running you from their own project, so a bare `scripts/transcribe.py` won't resolve. `--out` is the opposite: leave it relative to the user's working directory, so the archive lands in their project next to whatever they're working on rather than inside the skill install. Takes a URL or a file already on disk, and writes `transcript.txt` (readable, timestamped) and `transcript.json` (structured). Backend comes from the `TRANSCRIBER` env var — `whisper` (default, local, free), `deepgram`, or `azure`. It needs `yt-dlp` and