transcribe-video

name	transcribe-video
description	Transcribe YouTube/podcast/audio URL ra clean text. Auto-captions trước, fallback whisper-cpp + Silero VAD. Use trước khi /research video/audio source, hoặc user nói 'transcribe'/'get the transcript'/'get raw quotes from'.
context	fork

/transcribe-video URL [--out PATH] [--whisper] [--model MODEL] [--prompt TEXT] [--keep-vtt]

When to invoke

Always before any research/finding extraction from a video or audio source. Aggregator articles and same-cycle summaries hallucinate quotes — they conflate quotes from different talks by the same person. The verbatim transcript is the only trustable primary source.

Trigger conditions:

• /research is given a YouTube / podcast URL as the source
• User says "transcribe this," "get the transcript," "show me what they actually said"
• User pushes back on an aggregator-sourced report ("did they really say that?")
• A finding cites a video URL but quote provenance is [paraphrase] or [aggregator]

Skip conditions:

• Source is text (article, blog, gist) — go straight to WebFetch
• A _transcript-verbatim.md already exists for this URL in the findings folder

Output contract

The script writes two files:

File	Content
<out>/_transcript-verbatim.md	Cleaned, paragraph-reflowed transcript with frontmatter (source URL, voice, method, capture date)
<out>/_transcription-metadata.json	Build metadata: word count, method used, repetition zones detected, model + prompt, runtime

The verbatim file is the source-of-truth artifact. Findings, reports, and content angles cite only quotes that grep into this file. Everything else is paraphrase and must be source-tagged.

Two transcription paths

Path A — auto-captions (default)

yt-dlp --write-auto-subs --sub-langs en → fetch VTT → strip karaoke tags → dedupe consecutive lines → reflow into paragraphs.

• Speed: ~5 seconds for a 30-min video
• Cost: zero (no model)
• Coverage: complete (every spoken phrase)
• Quality: auto-caption typos on proper nouns (Claude Code → "lot code," OpenCode → "open claw," vibe coding → "vivibe coding," ChatGPT → "ChachiPT," etc.)
• Mitigation: the script applies a known-typo replacement table (references/typo-fixes.json); maintain it as new patterns surface

Path B — whisper-cpp (--whisper flag)

Download audio via yt-dlp → convert to 16kHz mono wav → run whisper-cli with Silero VAD + topic-seeded initial prompt.

• Speed: ~3-4 minutes for a 30-min talk on Apple Silicon (large-v3-turbo)
• Cost: zero (local model, ~1.6GB on disk)
• Coverage: can drop into degenerate repetition loops on certain content (whisper.cpp known issue)
• Quality: much better proper-noun recognition; Opus 4.7 / Codex 5.4 / Claude Code captured directly
• Mitigation: the script detects repetition loops (5+ identical consecutive phrases) and falls back to the auto-caption text in those zones, splicing the two sources by timestamp

When to use whisper:

• Auto-caption typos make quotes unusable for direct citation
• The video has accents / multiple speakers / technical jargon the auto-caption mangles
• User explicitly asks for "best quality transcript"

When NOT to use whisper:

• 30-second clip — auto-subs are fine
• User just wants a quick read of "what was said" — auto-subs more than enough

Recommended decision

Default to auto-subs. Re-run with --whisper if a finding's verbatim quote contains a typo'd proper noun the user will want to cite.

Initial prompt for whisper

--prompt seeds whisper's context with topic-relevant proper nouns. Without a prompt, whisper hallucinates novel spellings ("OpenClaw" / "CORTCO" / "JSON things"). Pull the prompt from:

1. --prompt "..." flag if user provided
2. The video title + description (yt-dlp --get-title --get-description)
3. Topic terms from the surrounding /research context

Always include in the prompt: speaker name(s), event name, and 5–10 likely proper nouns / acronyms.

Storage convention

Default output path: {vault}/knowledge/research/findings/{slug}/. Where slug is derived from --slug flag, the calling skill's findings dir, or <voice>-<topic> from yt-dlp metadata.

If the calling skill is /research, write to {vault}/knowledge/research/findings/{research-slug}/_transcript-verbatim.md so subsequent finding files in the same folder can [[wiki-link]] it.

Wiring into /research

When /research is invoked with a URL that's a YouTube watch link, podcast RSS item, or other audio/video source:

1. Run /transcribe-video first, write to the research's findings folder
2. Then proceed with normal research flow, but constrain quote-extraction to the verbatim file (NOT same-cycle aggregator articles)
3. Source-tag findings as [primary — verbatim] not [paraphrase] or [aggregator]
4. Aggregator articles can still inform structure (which topics matter) but never source quotes

Script

${CLAUDE_PLUGIN_ROOT}/scripts/transcribe-video.ts — TS+bun per global rule.

Direct invoke:

PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-$(ls -d ~/.claude/plugins/cache/brain-os-marketplace/brain-os/*/ 2>/dev/null | sort -V | tail -1)}"; PLUGIN_ROOT="${PLUGIN_ROOT%/}"
bun ${PLUGIN_ROOT}/scripts/transcribe-video.ts <URL> --out <DIR> [--whisper] [--prompt "...."]

Setup (one-time)

The script requires:

• yt-dlp (brew install yt-dlp)
• ffmpeg (brew install ffmpeg)
• For whisper path: whisper-cli (brew install whisper-cpp) + model files (auto-downloaded to ~/.cache/whisper-models/ on first whisper run)

If a dependency is missing, the script prints a single-line install command and exits with code 2. Don't try to auto-install — let the user run it.

Examples

# default — auto-subs, save to current research findings dir
bun scripts/transcribe-video.ts https://www.youtube.com/watch?v=96jN2OCOfLs --out knowledge/research/findings/karpathy-vibe-to-agentic

# high quality with whisper, with topic-seeded prompt
bun scripts/transcribe-video.ts https://www.youtube.com/watch?v=96jN2OCOfLs \
  --out knowledge/research/findings/karpathy-vibe-to-agentic \
  --whisper \
  --prompt "Andrej Karpathy at Sequoia AI Ascent 2026. Topics: vibe coding, agentic engineering, Software 1.0/2.0/3.0, Claude Code, OpenCode, Codex, NanoGPT, jaggedness, verifiability, Menugen, Nano Banana."

Failure modes & mitigations

Failure	Cause	Mitigation
Repetition loop in whisper output	Beam search degenerate state on long silence / repeated content	Script detects + splices auto-sub text into repetition zone
Auto-subs unavailable	Channel disabled subs / region-locked	Auto-fall back to whisper
URL is not YouTube	Generic audio file	Skip yt-dlp subs path, go straight to ffmpeg + whisper
Whisper missing model	First run	Auto-download large-v3-turbo (1.6GB) + Silero VAD (885KB) — happens once, ~30 sec on fast network
Wrong proper nouns in auto-subs	Caption upload defaults	Apply references/typo-fixes.json post-process; user can extend the file

Maintenance

references/typo-fixes.json — keep current. When a transcript has a new misheard proper noun, add it. The file is sorted by precedence (longer phrases first to avoid partial-match collisions).

When whisper-cli or yt-dlp behavior shifts (e.g. a new flag default), update the script. Keep the SKILL.md decision rules stable — those are the contract with calling skills.

Outcome log

Follow {vault}/skill-spec.md § 11. After the transcript is written (the skill's own outcome — not after the caller's research/findings step), append to {vault}/daily/skill-outcomes/transcribe-video.log:

{date} | transcribe-video | transcribe | ~/work/brain-os-plugin | {out}/_transcript-verbatim.md | commit:none | {result}

• result: pass — transcript delivered (either path) with no repetition-zone splices needed; partial — delivered but whisper repetition zones were spliced or new typo-fix entries had to be added; fail — both auto-subs and whisper paths failed, no transcript written.
• commit:none — transcripts land in findings dirs that are committed by the calling skill, not by this one.
• Optional: args="{url}" (enables replay), method=auto-subs|whisper.

If result != pass, auto-invoke /brain-os:improve transcribe-video.