Produce beat-synced motion-graphic videos with ak:motion-video

Use ak:motion-video to make a 1920×1080, 30 fps motion-graphic video where scene cuts land on the music's beats, reveals land on spoken words, and the audio mix is measured rather than guessed. The workflow is scripted end to end: facts, script, generated audio, forced alignment, beat grid, bar-spliced music, timeline and mix, HTML composition, lint and check, render, remux, social encode, and optional 1:1 or 9:16 editions. It renders through HyperFrames (HTML

• GSAP) and does not publish the result.

Choose the motion-video workflow

Use ak:motion-video when

• You need a launch, release, changelog, feature, or explainer video with voice-over, captions, SFX, and music synced to the beat.
• You want a restyle of an existing video as a separate project, or a partial rework such as new music, re-timed drops, a re-mix, a re-render, a social encode, or a square/vertical edition of one.
• You want one named style profile, or one base style with individual dimensions borrowed from other styles for particular scenes.
• Every on-screen fact can be traced to a primary source such as release notes, repository docs, or code.

Choose another workflow when

• You need generative text-to-video footage. Use ak:video.
• The composition is a React or Remotion project. Use ak:remotion.
• You have a general HyperFrames question or want a plain HTML composition without the beat-sync pipeline. Use ak:hyperframes.
• You need talking-head editing. This Skill does not cover it.
• You want a router across every installed motion or video Skill. Use ak:motion-graphics.

Meet the production bar

The workflow targets a real production piece, not a demo: every scene needs a focal point, depth, overlapping choreography, anticipation and follow-through, micro-motion through its holds, and a transition that carries into the next scene. Every scene passes a director's pass before the final render.

By default, a director subagent writes the creative direction and reviews the draft for taste, and an executor subagent runs the pipeline and builds the composition from that brief; the main session pins the outcome, owns the brief, and verifies the result. Pass --poor for a single session with no subagents, lean reads, a compact brief, and one review round when token spend matters more than the full director/executor split — it never skips a verification check.

Prepare the production route

Pin the outcome before any file is touched: topic and sources, style, duration, voice language, ending, and signature. English voice-over with Vietnamese captions is the default. The Skill asks only when the style or the facts source is unknown.

The full route also needs:

• The multix CLI on PATH, configured with its own credentials for Gemini TTS voice, ElevenLabs SFX, ElevenLabs Music composition plans, and ElevenLabs forced alignment. Credentials stay in multix's config; the Skill never prints or copies them into project files, plans, or reports.
• Node.js to run the project scripts and the pinned npx --yes hyperframes@0.8.77 commands.
• ffmpeg and ffprobe for the audio mix, remux, loudness and black-frame checks, and social encode.
• Python with NumPy and SciPy for the beat-grid and audio checks, and PyYAML for scripts/resolve-style.py.
• A writable project and enough storage for generated WAV and MP3 audio, snapshots, and 1080p MP4 renders.

Read Runtime adapters before assuming the same production surface everywhere.

Run the Skill

/ak:motion-video "Make a 60-second release video for our v3.4 notes in the glass-keynote style. Use only facts from CHANGELOG.md, English voice-over with Vietnamese captions, and stop after the master render; do not publish"

/ak:motion-video "Make a 60-second release video for our v3.4 notes in the glass-keynote style. Use only facts from CHANGELOG.md, English voice-over with Vietnamese captions, and stop after the master render; do not publish"

$ak:motion-video "Make a 60-second release video for our v3.4 notes in the glass-keynote style. Use only facts from CHANGELOG.md, English voice-over with Vietnamese captions, and stop after the master render; do not publish"

Add --poor to the same request for the lean, single-session mode.

Pick or mix a style

Every look is a style profile: a compact motion-design grammar for palette, typography, layout, material, texture, camera, transitions, pacing, beat response, UI choreography, diagrams, character motion, captions, and sound. references/styles/index.yaml lists about 60 of them. Brand names such as Apple, Nike, Linear, or NVIDIA are aliases that lead to a canonical motion language (for example, Apple-like leads to cinematic-product-launch); the output never uses a brand's logos, fonts, or assets.

• One style: name it, for example "comic style" for comic-multiverse.
• Mixed: keep one base and borrow individual dimensions for particular scenes. python scripts/resolve-style.py resolve data/style.json --script data/script.json resolves the mix into a per-scene table that goes into the plan.

Three styles also ship a detailed reference: comic-multiverse, glass-keynote, and cinematic-product-launch. The working projects behind their demos are not bundled; every new video starts from the Skill's composition skeleton plus the resolved style.

Follow the production stages

1. Pin the outcome. Topic, sources, base style and any layers, duration, voice language, ending, and signature are recorded.
2. Scaffold the project. The Skill creates assets/videos/<slug>/ and a plan under plans/, then copies the template scripts, example data files, hyperframes.json, and the composition skeleton. A finished edition is never overwritten; a restyle is a new project directory.
3. Write facts and script. Facts come only from primary sources, with a short evidence report before scripting. data/script.json holds the scenes and lines.
4. Generate and align audio. multix produces the voice-over, SFX, and music; ElevenLabs forced alignment maps spoken words to timing.
5. Fit the beat and arrange music. fit-beat-grid.py reports BPM and a per-bar table; bars are spliced so drops land where the script needs them, and verify-arrangement.py checks the splice.
6. Build the timeline and mix. build-timeline renders the ducked, loudness-normalized mix; measure-mix-balance.py reports music under speech.
7. Build and validate the composition. index.html follows the composition contract and the resolved style, then passes hyperframes lint, check, and scene snapshot review, plus the director's pass on every scene.
8. Render, remux, and encode. HyperFrames renders the video, the measured mix is remuxed back in, and an optional social encode produces a smaller file.
9. Add a square or vertical edition, when requested. A re-layout edition (scripts/reframe.py --ratio 1x1|9x16) reuses the master's voice-over, music, and mix and adds only ratio-specific CSS and its own render; a quick ffmpeg fit or crop is a faster, lower-fidelity alternative for previews.

Keep generation, spend, and publishing gated

Brand aliases select a motion language only. Confirm rights for fonts, images, music, voices, and any product visuals, and verify every on-screen claim against the collected sources.

Review outputs and evidence

A project typically contains:

assets/videos/<slug>/index.html
assets/videos/<slug>/hyperframes.json
assets/videos/<slug>/data/script.json
assets/videos/<slug>/assets/audio/{vo,sfx,music}/
assets/videos/<slug>/renders/<slug>.mp4
assets/videos/<slug>/renders/<slug>-social.mp4

The Skill reports measured numbers, not impressions. It is done only when:

• hyperframes lint and check pass and snapshots of every scene were reviewed, and every scene passed the director's pass.
• verify-arrangement.py reports every segment within ±2 ms, and measure-mix-balance.py shows the music about 4–6 dB under speech.
• ffprobe reports the expected frame size (1920×1080 for the master, or the requested 1:1/9:16 edition's own size), 30 fps, the expected duration, and AAC 48 kHz stereo; loudness is about −14 LUFS with peak at or below −1 dBFS, and blackdetect finds nothing outside fades.
• build-timeline printed no clip, overlap, or missing-word warnings.
• Each requested 1:1 or 9:16 edition passes the same lint, check, ffprobe, loudness, and blackdetect checks as the master.
• On-screen facts trace to the sources collected at the start.

Troubleshoot safely

Continue with the Marketing Kit overview, Projects, artifacts, and checkpoints, or Installing kits.