footage-ingest

The 'card is in' step: after a recording session, finds the camera or mic card (DJI Osmo/Action/Mic, GoPro, Sony, any DCIM card), copies every clip with size verification, checks audio length == video

SKILL.md

GitHub

Footage ingest: "the card is in"

Step 3 of the pipeline: idea -> script (reel-ideation) -> record -> "the card is in" -> edit-style -> trial-post -> reel-scorecard.

The creator records, plugs the card (or the camera over USB) into the computer, and says "DJI is in". This skill takes it from there and ends with a TAKE-MAP.md that hands the editor one folder per reel with the chosen clips.

Requirements

Claude Code on the computer the card is plugged into (a chat app cannot read a card). ffmpeg + ffprobe, Python 3.9+, a local Whisper (use-local-whisper skill). Check: ffprobe -version && python3 --version.

Config (ask once, save to ingest/config.json in the creator's content folder)

key meaning default
dest_root where sessions go $HOME/reels
mirror_camera flip camera footage horizontally (selfie view). Ask the creator: many prefer how they look in a mirror ask
fps conform frame rate 30
min_free_gb free disk needed before copying and rendering 40
record_sheet the sheet the slates refer to latest RECORD-SHEET.md

Steps

  1. Before: if the creator is still recording, do nothing heavy (no renders, no transcription): an overloaded machine can drop mic audio. Start when they say the card is in.
  2. Find the card. python3 scripts/ingest.py find lists every mounted folder with clips (DJI cameras write DCIM/DJI_###/DJI_<YYYYMMDDHHMMSS>_<NNNN>_D.MP4; DJI Mic transmitters mount as their own volume with WAV backups). Pick the folder whose clip times match the session. More than one card: ingest each.
  3. Check disk. At least min_free_gb free (renders and swap need it). If not, list what could be cleared and ask. Never delete anything on the card, and never delete raws on the computer to make room.
  4. Copy + verify. python3 scripts/ingest.py copy --dest <dest_root>/record-<date> --src <card folder>. It copies MP4/MOV + mic WAVs (proxies like .LRF stay on the card), size-checks every copy, re-runs safely, and writes manifest.json + durations.tsv. Exit 1 = a failed copy or an A/V mismatch.
  5. Audio == video, every clip. Any AV MISMATCH or NO AUDIO line: tell the creator at once, name the clips, and ask for a re-take while the set is still up. If a mic WAV backup covers that clip, note it as the audio source instead.
  6. Eject. Only when step 3 printed 0 failed copies: sh scripts/eject.sh <card folder>. From here on work only from the copies.
  7. Transcribe, one clip at a time. sh scripts/transcribe_all.sh <dest>/raw/<folder> <dest>/transcripts. One heavy job at a time, nice, 4 threads: parallel 4K decodes plus Whisper can exhaust memory.
  8. Sort. Read every transcript next to the RECORD-SHEET. Mark each clip:
    • which card and part it is (by the spoken slate, else by matching the words to the script): R6 hook B, R6 body, R6 CTA IG, R6 CTA bio;
    • FLUB (restart, trailing off), RETAKE OF <clip>;
    • NOT FOR REELS (other projects, tests, room tone): leave it in raw, do not edit it;
    • claims check: any spoken line that is false, outdated or overclaims ("scrape the entire internet", a number nobody measured) is flagged; the editor cuts it or the take is rejected.
  9. Pick the best take of every hook, body and ending: complete, clean, the last good take of a repeated line, no false claims. Write the reason for every rejection. List what the sheet asked for that was NOT recorded.
  10. Conform the picks (only the picks): python3 scripts/conform.py <dest>/conformed <clip>... [--mirror]. Mirror happens here, on the raw camera clip, before any caption, logo or screen capture is added (those are never mirrored). After mirroring, look at one frame per setup: text behind the creator now reads backwards (shirt logos, signs, screens); blur, crop or flag it. Phone screen recordings are never mirrored.
  11. Write TAKE-MAP.md (templates/TAKE-MAP.md): the map table, the SHIP list (one line per reel: hook clip > body clips > IG ending / bio ending), REJECTED with reasons, NOT RECORDED, and open questions. Then hand each SHIP line to edit-style, at most 2 edits rendering at once on one laptop.

Output

record-<date>/
  raw/<card folder>/       untouched copies (never edited, never deleted)
  manifest.json  durations.tsv
  transcripts/<clip>.words.json / .txt / .srt
  conformed/<clip>.mp4 + conform-manifest.jsonl
  TAKE-MAP.md              what ships, what was rejected and why, what is missing
  LOG.md                   written as you go (a long ingest outlives one agent)

Failure handling

problem do
no card found ask the creator to plug it in or connect the camera in USB storage mode; INGEST_CARD_GLOBS for odd mount points
copy size mismatch twice do not eject; re-seat the card, run copy again (it skips good files)
A/V mismatch name the clips, ask for a re-take now; check the mic backup WAV
disk under the limit stop before transcribing; list space to clear, ask
a slate is missing match on the words; if two scripts fit, say so in TAKE-MAP open questions
the agent gets cut off mid-ingest everything is re-runnable: copy skips verified files, transcription skips done clips, LOG.md says where it stopped