Dji Sync

DJI Sync

This skill starts when the recording stops and owns the card until it is empty. It does not edit, trim, colour-grade or re-encode video. It swaps one audio track and gets out. Hands off to shorts-production (encoding, Drive, ClickUp coding and scheduling), descript-projects (getting footage in), video-hooks and generate-social (hooks, scripts, captions), and the vault's own CLAUDE.md ingest contract (filed voicenote to wiki pages).

Two paths off the same card

A take is either half of a reel or a thought recorded alone. They share the card detection and the archive-and-wipe contract and nothing else.

Ask which one it is. Do not infer it silently. The speaker says it in plain language ("that's a voice memo", "pair this with the reel") and that always wins. A wrong guess either burns 20 minutes of thinking or destroys the audio half of a reel. When he has not said: no clip in ~/Downloads near the take's timestamp means solo; a long take that VAD reports as mostly silence means solo, a walk with pauses; a take within seconds of a clip's length means reel. Run dji_note.py --dry-run when unsure. It decides and reports without writing anything.

The rule that ranks above the rest

The video is the timeline. The clip's duration, frame timing and metadata are the truth. The DJI take is a raw ingredient: it gets trimmed, shifted and padded to fit the video, never the other way round. Output duration always equals input duration. If that is not true, the run failed.

Normally the mic is started first and stopped last, so the take fully contains the clip and both ends get trimmed off. That is the expected case, not a guarantee. See "When the take runs short".

Never match on timestamps

The transmitter's clock and the phone's clock do not agree, and the drift is not constant. On the first real pair the DJI filename stamped 190732 while the clip stamped 19:07:49, implying 17 s, and the true measured offset was 16.004 s. A second take on the same card was stamped 90 s away from where its audio actually sat.

Filename stamps are worth using to order takes and nothing else. The waveform decides.

How the match is decided

The pipeline aligns once globally with GCC-PHAT, then re-measures the offset independently in every 2-second window: a genuine match holds the same offset everywhere, a false one wanders. Alignment is sub-sample and drift is corrected when it matters. The measured evidence behind the method (why envelope correlation and peak height fail, the drift numbers) is in references/matching-method.md.

Accept requires all of: ≥ 80 % of windows within 1.0 ms of the median, each with peak-to-sidelobe ≥ 5; a ≥ 15 percentage-point margin over the runner-up take; at least 2 usable windows. Miss any and it refuses and writes nothing. A wrong take burned silently is far more expensive than a refusal: you find it in the edit, an hour later. Do not loosen these to force a match. Pass the take explicitly instead: dji_sync.py clip.MOV --take /path/to/TX00_MIC005.wav.

When the take runs short

The very first real pair covered 20.29 s of a 35.84 s clip: the transmitter stopped before the phone did. Assume it will happen again. Uncovered stretches are filled from the phone's own audio, RMS-matched to the lav so the level does not jump, with a 40 ms crossfade at the seam. The run prints a loud WARNING with the head and tail gap. --silence-gaps leaves them silent instead, which makes the gap obvious in an edit at the cost of throwing away usable audio.

Always read the coverage line. A file at 57 % coverage is a usable rescue, not a clean take.

What "no loss" means here

• Video is never re-encoded. -c:v copy, and the burn is rejected unless the output's video stream MD5 equals the source's. Bit-identical or it does not ship.
• Audio is lossless. ALAC s32p in the .MOV. Measured round-trip residual −143.3 dBFS, max 1.19e-7, which is float32's own epsilon rather than any codec loss.
• The only intentional changes to the lav samples are the sub-sample shift and the seam crossfade.

--codec pcm for older NLEs, --codec aac for a small upload copy. Both a step down for no gain in an edit.

Metadata and what survives

-map_metadata 0 -movflags use_metadata_tags, then exiftool restores the Apple keys ffmpeg drops, then the file's mtime and creation date are reset to the original's. Verified preserved: rotation (90°, both as the tag and the display matrix), CreationDate with its timezone, GPS, Make/Model/Software, and the four mebx Core Media Metadata tracks with their handler names. macOS parses the result natively.

Known cosmetic wart: ffprobe reports the data tracks' codec_tag_string as stts rather than mebx after the remux. The handlers and payloads survive and macOS reads the file. If a tool ever chokes on those tracks, drop them rather than shipping mangled ones.

The solo path: silence goes before whisper sees it

Silence stripping is a correctness requirement, not a size optimisation: fed a mostly-silent take, whisper loops on one sentence and the real content never appears. Silero VAD decides what is speech; a fixed dB gate needs per-recording tuning and is only a cross-check. The measured evidence for both claims is in references/matching-method.md.

Everything else is defaults with a reason: 0.25 s of padding either side so words are not clipped, joins de-clicked with a 10 ms fade, and silences shorter than 400 ms left in.

The loop guard, which is the real safety net

The transcript is checked before anything is written or deleted. If one normalised sentence accounts for 6 or more lines and 40% of the transcript, it is whisper looping, not the speaker repeating himself. The run refuses: no note, no wipe, card untouched, and it says to try a lower --vad-threshold. Verified against both transcripts above: the trap trips at 38 of 42 lines, the real one passes at 1 of 11.

Where the note lands

📥 Raw/Voicenotes/YYYY-MM-DD <Title>.md in the vault, with the take id, both durations and the audio path in frontmatter.

Raw is immutable and the wiki is a separate act. The script files the transcript and stops. Turning it into 🧠 Wiki/Coaching/2026 North Star.md, the project pages, index.md and log.md is the vault's ingest, which its own CLAUDE.md says runs one source at a time with the speaker in the loop. Summarise the take back and ask what to emphasise before writing a single wiki page. Never quietly rewrite the North Star from a walk.

Pass --title once the transcript has been read. Without it the note is filed under a provisional DJI note HHMM and has to be renamed during the ingest.

What the wipe destroys here

The reel path archives every take in full. The solo path does not. It archives the speech-only Opus and deletes the source wav, which is irreversible: the 95% that was silence is gone, and so is the untouched audio of the 5% that was not. The transcript must pass the loop guard before anything is deleted, so the ordering still holds. When the raw take matters, --keep-original archives the full wav through the same checksummed path the reel side uses.

Running it

D=<skill-dir>/scripts
python3 "$D/dji_sync.py"                     # newest clip in ~/Downloads, detect card, burn, wipe
python3 "$D/dji_sync.py" --dry-run           # decide and report, touch nothing
python3 "$D/dji_sync.py" clip1.MOV clip2.MOV # several clips, each take used at most once
python3 "$D/dji_note.py" --dry-run --no-note # cut, transcribe, print, write nothing
python3 "$D/dji_note.py" --title "..."       # file the note, then wipe the card

--no-clean keeps the card as it is on either path.

Reel flags: --take to force the pairing, --keep-scratch to retain the phone audio as a second track, --silence-gaps, --codec, --archive DIR, --card PATH, --output PATH. Note flags: --title, --vad-threshold (0.5), --min-silence-ms (400), --pad (0.25), --bitrate (24k), --lang (auto, so French and English both work), --keep-original, --no-note, --vault DIR.

Reel output lands beside the source as NAME-dji.MOV. The original is never modified, so a bad run costs nothing but disk. Note output lands in the day's archive folder and the vault.

What it runs on

Both paths need ffmpeg, ffprobe and numpy. The solo path also needs whisper-cli and whisper-vad-speech-segments (brew install whisper-cpp), plus two models already on this machine because VoiceInk downloaded them:

• ~/Library/Application Support/com.prakashjoshipax.VoiceInk/WhisperModels/ggml-large-v3-turbo-q5_0.bin
• /Applications/VoiceInk.app/Contents/Resources/ggml-silero-v5.1.2.bin

Nothing is downloaded and nothing is sent anywhere: it is all local and free. This borrows another app's files, so if VoiceInk is ever removed, copy both models somewhere stable and point --model and --vad-model at them rather than re-downloading. A 20-minute take costs about 26 seconds end to end on the M4 Pro.

The card wipe, and the trap in it

Order is fixed and not negotiable: burn → verify → archive → checksum → delete. Every take on the card is copied to ~/Movies/dji-archive/YYYY-MM-DD/ and each copy's SHA-256 is compared against the original before anything is removed. A take whose checksum mismatches stays on the card. If any clip in the run failed to verify, the card is left completely alone.

Dragging takes to the Trash on a removable volume never frees the space. They move to a hidden .Trashes folder on the card and sit there. The first run found 31.5 MB of takes "deleted" that way. The wipe clears .Trashes explicitly, which is most of the point of automating this.

Verification

• New failure modes from this run appended to Learned Patterns

Learned patterns

This skill appends new failure modes to its own pattern list after each run. The pattern list lives in references/learned-patterns.md. Read it before a run. If this run surfaced a failure mode not already listed, append it there with today's date before finishing. Keep the measured number, not the impression.