Capafy logo Capafy
开始使用
Footage Librarian

Footage Librarian

by: SeanWithAI
Drop raw videos, photos and voice memos into one folder and Claude catalogs them. Each file gets an ID, a frame sheet and a transcript made on your Mac. Claude looks at every item once and writes up a summary, who's in it, quotable moments with timestamps, what it's good for and anything that shouldn't be posted. The result is one searchable CATALOG.md, so you can find a clip without re-watching your footage. Needs macOS on Apple Silicon, ffmpeg and mlx-whisper.
#视频#社交媒体
评分
需要更多评价
订单数
0
如何使用
下载

Footage Librarian

You film a lot and can never find the clip you need. Footage Librarian turns a folder of raw videos, photos and voice memos into a catalog you can search, so each clip is watched once and never again.

How it works

  1. Drop new files into a raw/ folder and tell Claude "new raw files".
  2. Each file gets an ID (V0001 video, P0001 photo, A0001 audio) and moves into raw/videos/, raw/photos/ or raw/audio/. Originals are never re-encoded, trimmed or deleted.
  3. A script builds what Claude needs to look at it once: a sheet of 12 timed frames for videos, a preview for photos, a waveform for audio, and a transcript made on your Mac with mlx-whisper.
  4. Cheaper subagents (Haiku for silent clips and photos, Sonnet for anything with speech) describe the items in parallel batches.
  5. Each item is renamed to ID_date_slug and added to raw/CATALOG.md.

What each entry holds

  • A 2–3 sentence summary
  • Type: work, daily life, failure, opinion, how-to or b-roll
  • Who is in it (only people you list; everyone else is described, never guessed)
  • Location, project and tags
  • Quotable moments with timestamps taken from the transcript or frame times
  • Quality notes: light, focus, audio level, shake
  • What it's good for: hook, talking head, b-roll, screen demo, thumbnail, and the 9:16 crop it needs
  • Privacy flags: children, addresses, car plates, readable screens, strangers' faces

Transcripts you can trust

Whisper makes up text over silence ("Thank you for watching"). Footage Librarian measures the audio, re-runs quiet clips boosted, and says so when a clip has no usable speech, so you don't build a video on a line nobody said.

Also handles

  • Photo bursts, grouped as one item with a contact sheet
  • Voice memos named after a street (flagged as an address)
  • Duplicate files, skipped by content
  • Files dropped into the wrong subfolder

Setup

On first run Claude creates raw/ and a short profile.md where you list the people who appear in your footage, your projects and your privacy rules.

Requirements

  • macOS on Apple Silicon
  • Claude Code
  • ffmpeg (brew install ffmpeg)
  • mlx-whisper (pip install mlx-whisper). The first transcript downloads the whisper-large-v3-turbo model, about 1.6 GB.

Transcription runs on your Mac. Your Claude session sees the frame sheets and transcripts it describes, under Anthropic's terms.


View more