board-sync

Animate explainer boards to the presenter's voice. Takes the board HTMLs from Claude Design (04-boards), the final-cut WAV and its words.tsv, the marked script and the board brief; makes every element

SKILL.md

GitHub

board-sync

Step 5 of the explainer pipeline: the boards from Claude Design plus the final-cut audio become a board video where every element appears on the word that names it, and a timeline the editor opens in DaVinci Resolve or Premiere.

What it needs (the video's folder)

02-cut\<name> - final cut.wav            the final audio (the only sound in the video)
02-cut\<name> - final cut words.tsv      word timings from your editor (if missing, board-sync transcribes the WAV)
02-cut\<name> - final cut.mp4            optional: the 4K cut, only to name it in the timeline
03-board-brief\script-marked.md          the script with [▸ SLIDE n · title] and [CAMERA] markers
03-board-brief\brief.md                  the board brief (builds, anchors, numbers' roles, fillers)
04-boards\NN-Name.html + assets\         one HTML per board from Claude Design, with its assets folder

Setup (once per computer)

Windows, in PowerShell inside this skill's folder:

powershell -ExecutionPolicy Bypass -File .\setup.ps1

Mac: bash setup.sh. Both are safe to run again. They install uv, a Python 3.12 environment with the packages, ffmpeg, and check for Edge or Chrome (Playwright's own Chromium is downloaded only if neither is there). The first transcription, if one is ever needed, downloads the speech model (about 1.6 GB). The setup ends by printing the exact run command for this computer.

Run (from the video's folder)

& "<skill>\.venv\Scripts\python.exe" "<skill>\board_sync.py" --boards "04-boards" --wav "02-cut\<name> - final cut.wav" --words "02-cut\<name> - final cut words.tsv" --script "03-board-brief\script-marked.md" --brief "03-board-brief\brief.md" --out "05-board-sync\out"

On a Mac the same with <skill>/.venv/bin/python and forward slashes. Quote every path. Never end a quoted folder path with a backslash ("04-boards", not "04-boards\").

  • --out can be left out: it defaults to 05-board-sync\out next to 04-boards.
  • --only 4,5 redoes only those boards after a fix; --stage words|windows|brief|units|choreo|render|qa|timeline|preview runs one stage (each stage reads and writes --out, so stages can be re-run alone).
  • --cut "<file>.mp4" names the 4K cut in the timeline (otherwise it names 02-cut\<name> - final cut.mp4).
  • The background is kept unless the editor passes --mask <file> (the cutout luma matte from the cutout step). Only then does the timeline get the matte track above the cam; --camera-keyed says the full-frame camera stretches are keyed too.
  • --check shows what the computer has (Python, ffmpeg, browser, packages) and which stages are ready.

What comes out (05-board-sync\out\)

File What it is
board-track.mp4 every board at its window, black on camera stretches, exactly as long as the WAV, no audio
boards\NN-Name.mp4 one clip per board
timeline.fcpxml DaVinci Resolve: File, Import, Timeline
timeline-premiere.xml Premiere: File, Import
preview.mp4 the board track with the WAV, for the watch-through only
qa-report.md, sheets\NN.jpg what to fix, per board, with contact sheets
words.json, windows.json, sync-manifest.md word timings, where each board sits, what lands when

The timeline: V1 the board track at 200 %, V2 the 4K camera cut at every window (32 % in the bottom-right zone, 80 px from the right and bottom edges, on boards; full frame on camera stretches; its own audio off), A1 the WAV, no effects. python scripts/timeline_check.py 05-board-sync\out prints every clip's in/out, scale and position.

The QA loop (do this before going on)

Read qa-report.md. Fix and re-run until it is clean:

  • An element lands before its word, or more than about 2 s after: a timing fix in board-sync (usually the anchor words in brief.md), then re-run that board with --only N.
  • More than about 3 s of speech with nothing new landing: board-sync adds a filler (the presenter's exact words). If it says "uncovered > 15 s", a board is missing: back to steps 3 and 4 for that board only.
  • A price sticker on a count ("5 out of 13" showing "$13"): fix the number's role in brief.md, re-run.
  • Board copy that contradicts what the presenter says: fix the board in Claude Design (never the presenter's words), download that board again, re-run.
  • Anything in the camera zone: fix the board in Claude Design.

Then watch preview.mp4 from start to end. Judge by watching. Screenshot anything static, empty or out of sync and fix it.

Before the first render, look at the boundary table the windows stage prints (each board's start and end with the words around the start). Every board should open on the first words of its block. A board flagged under 40 % matched means its script lines are not what the presenter said in the final cut: fix script-marked.md, not the timing.

The rules (all enforced in code; see references/rules.md)

  • Land on the spoken word, never before it, never more than 2 s after.
  • Named things appear when they are named. A part never appears without its container.
  • Structure first: boxes, rails and outlines are there as the board opens; no board opens empty.
  • Money and percent stickers only on money words; a count never gets a price sticker.
  • Nothing static for more than 3 to 4 s while the presenter talks: fillers are the presenter's exact words.
  • Nothing moves after it lands: no zooms, no push-ins, no pulses, no camera moves.
  • No music, no sound effects. The WAV is the only sound; no limiter on the voice.
  • The camera zone (bottom right, 615 x 430 px at 1080p) stays empty on every board.
  • Board copy that contradicts the presenter's words is flagged, never silently changed.
  • Ad-libs are never dropped: every spoken word stays in the window it falls in.

When something goes wrong

See references/troubleshooting.md: no browser found, ffmpeg missing, words.tsv missing, a board under 40 % match, fonts not loading, Windows path quoting, long renders.

Notes for Claude running this skill

  • Exporting the full board video (player with camera, clips and voice): follow the export rules in the html-to-mp4 skill. Every clip seeks on every frame, preflight 90 frames first, detached MKV workers, keyframe every 15 frames, TS join, then verify frame count, keyframe gap and motion before reporting. The AI Video Editing System project has this as 05-board-sync/export.sh.

  • Run the whole thing with one command; re-run single stages only to fix something. Exit code 0 means every stage that exists ran; QA findings are in the report, not in the exit code.

  • Contracts between stages (file names, JSON shapes) are in DESIGN.md. scripts/common.py holds the shared helpers.

  • The test fixture is built by tests/make_fixture.py <v2 project> <deck html export> into fixtures/ (generated data, not shipped).