Video production, resumable and cheap

Why this playbook exists

Renders are slow and media is opaque: Claude cannot watch a video, and ffmpeg progress or media bytes in the conversation are pure waste. Treat a film as data (a segment list) plus a deterministic build. Then any interruption resumes from the last finished segment, and any change re-renders only what it touches.

First: find the pipeline that already exists

• Look for docs/ARCHITECTURE.md, Video/pipeline/, build/, *.ffmeta, concat_*.txt, parts/, verify.json, parts_results.json. If they exist the project already has a pipeline: read the pipeline section of ARCHITECTURE.md and reuse it. Do not write a second one.
• On "continue": python <skill>/scripts/parts_status.py build/parts --expect segments.txt lists what exists, what is missing, and durations. Resume at the first missing or failed segment. Verify finished parts with ffprobe, never by re-rendering.

Structure a film as data

• One segment list (segments.json or the project's equivalent) with one entry per segment: id, source clip, in and out points, title card text (Arabic and English), lower-third, subtitle cue. This file is the film; everything else derives from it.
• build/parts/<id>.mp4 per segment, build/concat_<chapter>.txt for ffmpeg -f concat -safe 0, build/chapters.ffmeta for chapter metadata, <name>.srt beside each final mp4.
• Graphics (title cards, lower-thirds, outro) come from the project's DESIGN.md tokens. Render each once as PNG or HTML and reuse it; never restyle per segment.
• Arabic text: shape with arabic_reshaper and python-bidi when drawing with PIL. A full ffmpeg build (for example the gyan.dev full build) includes libfribidi and harfbuzz for drawtext; test one frame before rendering a chapter. Fonts follow the design system (Cairo for display, Tajawal for text).

Render rules

• Quiet ffmpeg: -hide_banner -loglevel error -nostats -y. Run long renders in the background with output to a log and check tail -n 5 render.log; never stream progress into chat.
• Facts, not eyes: ffprobe -v error -show_entries format=duration:stream=width,height,r_frame_rate -of default=nw=1 file.mp4. When a visual check is truly needed, one contact sheet (-vf "fps=1/10,scale=320:-1,tile=6x4") replaces many screenshots.
• Idempotent: skip a part whose output exists and is newer than its inputs; keep parts_results.json with status and duration per part.
• Subtitles: generate the SRT from the segment timings; burn in with the subtitles= filter only for the final export and keep the sidecar .srt too.
• Portrait and landscape variants share parts: crop or scale at concat time, not by re-recording.
• Never read .mp4 or .png bytes, never paste ffmpeg output; report durations and file names only.

Recording new footage

Use the project's recorder script (browser automation driven by a step list, usually documented in ARCHITECTURE.md with its environment variables). Script the navigation as steps, record per chapter, and take the screen list from the system inventory. Do not drive the browser manually screen by screen through MCP tools to record; that costs a screenshot per step and is not repeatable.