Short-form video analysis tool — crawl, transcribe, and visually analyze YouTube Shorts, Instagram Reels, and TikTok

Crawl, transcribe, and visually analyze YouTube Shorts, Instagram Reels, and TikTok videos into structured data.

inspect starts a small local web app for one clip and opens it in the browser. The video player is the single source of truth: a waveform (ffmpeg peaks, cached) with a click-to-seek playhead, a keyframe filmstrip, and the transcript all seek the player and highlight as it plays. Set IN/OUT markers on the waveform and export the trimmed window as SRT. Needs the downloaded video file on disk. (Design ported from arkiv's live inspector.) export --format html remains the offline multi-clip bundle; view is the read-only browsing server.

Whisper large-v3 locks onto the language it detects in the opening window and, on long files, "translates" later speech of the other language back into the locked one — a code-switching interview (Chinese host + English guest) comes out with the guest's English mangled into garbled Chinese. It is a long-form drift, not a bad audio issue: the same passage transcribes perfectly when sliced out on its own.

Fix — force per-chunk language re-detection:

multilingual alone is not enough — it needs a short chunk_length (~15s) so each chunk re-detects. Verified on a 40-min ZH-host/EN-guest interview: latin-char recovery 56% → 90%. Leave OFF for single-language short-form (per-chunk detection adds cost). Other levers: WHISPER_LANGUAGE=en (force one language), WHISPER_TASK=translate (force English output).

Reel Scout's pipeline gets clean input into a model. The reverse-decode prompt pack in prompts/ is the analysis brain you point at that input — to reverse-engineer why a short-form video works and extract a transferable structure, with anti-hallucination guardrails (observation vs. inference, cite the timestamp). Open (MIT). See prompts/README.md.

Video extraction techniques (captions-first transcription, duration-aware frame budgeting, time-range focus, on-screen-text resolution bump) are adapted from claude-video (MIT). See NOTICE.

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Filter files by name, interpreter, ABI, and platform.