What does designos-beat-edit-video do?
Purpose: marketing the design:os studio app. Every video shows the real app doing something a viewer wants (one feature, one clear promise, one call to action). The app is the hero, not the editing trick.
Not a fixed template. Every video must be different from every earlier one: new idea, new feature or task shown, new footage, new structure, new palette and background, new motion language, new music. Creativity has to stand out in the first 3 seconds. Before planning, read the earlier briefs in the Project (orca-briefs/*designos*) and pick choices that none of them used; write the choices into the new brief so the next video can avoid them. (engine/ideate.py, engine/music.py, references/creative-brief.md help with this.) spec/design-in-4-4.spec.json is only one worked example (its footage is old: never reuse it, always record new footage).
One spec JSON drives everything: video (render.py), audio (audio.py), checks (beatlib.validate, qa.py). A new video = a new spec + new raw takes. No code changes.
Batch: python engine/ideate.py batch --count 30 --seed N --out batch/ makes N different plans (style, bpm, concept, palette, structure, edits, length, feature) plus manifest.json. See references/chatgpt-instructions.md.
Hard rule
All footage comes from the real design:os studio app (Openscreen --window "design:os studio"), recorded fresh for each video. No mock-ups, no other app, no old takes.
Pipeline
python engine/ideate.py plan --edits N --seed <new> --out plan.json-> concept, music style + bpm, palette, structure, where each edit lands. Rewrite the copy yourself. After recording:ideate.py build plan.json times.json --out spec.json. Music: 5 generative styles (house, lofi, trap, synthwave, cinematic), set inspec.music.- Record real footage of the app (Openscreen,
--window "design:os studio", cursor editable-overlay). Rules: click, pause about 0.8 s, then type. Show the whole app. Human-like cursor. Record longer than the action. - Fill the spec (
spec/design-in-4-4.spec.jsonis the working example; fields inreferences/spec-fields.md). Per edit you only need raw-footage times:tc(cursor starts to click the sidebar),hit(screen changes),td(double-click on the text),ts/te(first/last typed character), plusT= the bar line (video seconds) where the edit lands. python engine/run.py validate spec.json- fails if an edit is off a bar line, typing is off the 16th grid, the click-to-text pause is too short, or scenes do not tile the timeline.python engine/run.py all spec.json --root <dir with raw takes> --out <dir>->final.mp4(1920x1080, 30 fps, AAC) and runsqa.py. Partial:render F N(scenes),audio,mux,qa.- Look at 3-4 frames and the contact sheet before delivering. Nam posts himself; never post.
Guardrails (Marketing HQ)
FREE, say "source-available" (never "open-source"), no "Figma killer", "Developer view", only shipped features, every claim backed by real footage. The video is about the product, not about the making.
Why it works
Footage time is mapped to video time by anchors (beatlib.Timeline.anchors), so a click lands on the bar line
whatever speed it was recorded at. 120 BPM: beat 0.5 s, bar 2 s, one typed character = 0.125 s.
Verified
Scene F re-rendered from the spec is frame-identical (PSNR inf) to the original; the audio WAV is byte-identical.
Default peak 0.89 gave +0.7 dBFS true peak on the sample; for new videos set audio.peak to about 0.8.
Requirements
python3, numpy, Pillow, ffmpeg/ffprobe; macOS fonts (HelveticaNeue.ttc, SFNSMono.ttf) or edit fonts in the spec.
legacy/design-in-4-4/ keeps the original scripts (tl.py, make8.py, audio9.py, cursor2.py) for reference.
Prompt Motion: use it on every video
The Prompt Motion Bridge tools (Prompt_Motion_Bridge__recommend_motion_references, search_motion_references, get_motion_reference) are the source of motion ideas. For every video:
- Describe each scene (hook, chapter cards, transitions, callouts, captions, stat/recap, end card) and ask for references; search with fresh keywords each time (kinetic type, mask reveal, morph, parallax, split-screen, glitch, liquid, stagger, counter, 3D tilt, ...).
- Pick at least 5 different references per video and build them in code (PIL/numpy/ffmpeg or Remotion). Note the reference ids in the brief and do not reuse the same set as the previous video.
- Bright, lots of animation, but the app footage stays the hero and readable.
Long-form tutorial (3-5 min, Vietnamese voiceover, YouTube)
First one approved by Nam on 2026-11-16 (final_v3.mp4, 4:21: 6-step process from a .fig template to desktop+mobile screens). Treat it only as an example: the next long video needs a different topic, footage, structure and look.
Starter code: longform/ in this skill (also on the Mac at longform/): README.md, script2.py, gen_tts.py, lib.py, build.py, audio.py. Copy it into a new workspace and change it: chapters, layout, card and caption design, background, palette, motion (from Prompt Motion), music style and tempo. Never ship the same look twice.
Ideas: each video picks its own angle and format, e.g. a challenge against a timer, before/after makeover, one feature deep-dive, "3 mistakes beginners make", redesign a real-looking screen, build a mini design system, mobile-first, speed-run, Q&A. Every chapter must show something the viewer wants; a chapter that is only "a mistake I made" was rejected as irrelevant. Shipped features only; check in the app before promising one (Play preview at top-right ▷ works and looks great).
Footage: record fresh with Openscreen for every video (different template or file, different screens, different edits). Openscreen records only the app window, not native macOS dialogs: show the file picker as a text callout, never a mock-up. Select text layers in Layers, press Return, check the text is highlighted, then type; typing into an unfocused field fires shortcuts. Before the voiceover claims an edit, check the app state again (an undo once silently reverted a button text).
Voice (approved by Nam): VieNeu-TTS v3 Turbo, preset "Hải Đăng". Runs on the Mac (.venv-voice/bin/python, gen_tts.py), 48 kHz, ~0.3x real time. Do not clone Jang's voice without his written OK.
- Generate sentence by sentence with 0.18 s of exact zeros between sentences; captions are timed from those gaps, so display text and TTS text must have the same sentence count.
- Keep two texts: display (captions, correct spelling) and TTS (respelled). Approved: "Design OS Studio" -> "Đi-giai Âu-ét Studio" (keep "Studio" in English), "Figma" -> "Phích-ma", ".fig" -> "phích", "megabyte" -> "mê-ga-bai". Never ALL CAPS "DESIGN OS" (spelled letter by letter).
- Nam rejects wrong English pronunciation: speak Vietnamese for UI words (bảng lớp, vùng vẽ, bảng thuộc tính, phím xuống dòng, phím thoát, màu nền); English UI strings go only in callouts.
- New respelling: try 5-7 variants, check them with faster-whisper (vi and en), send Nam a short sample WAV, and only rebuild after his OK.
Picture rules: 1920x1080, 30 fps. Openscreen raw is only 1600 px wide, so keep camera zoom at or below ~1.6x (higher looks soft on YouTube). Fit all text inside its card (auto-shrink) and check recap/stat frames for overflow. End card shows the full URL https://jangtrinh.github.io/design-os-studio/ large for 8-10 s plus "Source-available · Not affiliated with Figma".
Delivery: export 1080p plus a 4K upscale (lanczos, crf 16); send the 4K for YouTube, only after ffprobe reads its full duration. Tell Nam to keep the upload private until YouTube shows 2160p. Give one clear title suggestion plus a description with chapter timestamps, the link, the template credit and "Not affiliated with Figma". Never post; Jang approves last.
Reference files in this bundle
references/openscreen.md+tools/record.sh: exact Openscreen command, where takes land, cursor.json format, screen mapping.references/voice.md+longform/gen_tts.py+longform/check_voice.py: approved voice, install, respelling table, Whisper check.references/prompt-motion.md: how to use the Prompt Motion tools, references already used, fallback without the MCP.references/chatgpt-instructions.md: handoff for ChatGPT/ORCA/Codex, now with the long-form mode.longform/: starter engine for the long video (ROOT=<workspace> python build.py info|agenda|<chapter>,python audio.py).