Opus 5.5 Motion Graphics · Motion OS and a Two-Level Launch Video Workflow

Quick Summary

• Jason Lee divides the build into two levels. Level 1 is a 45-second launch film made from interface elements, typography, chat bubbles, phone screens and generated profile photos. Level 2 combines the same motion system with cinematic AI footage of a consistent founder in a consistent office.
• Both levels begin with a reference film. The Rene launch supplies the motion language for Level 1, while Anthropic's Claude Tag launch supplies the structure for Level 2. The goal is to study pacing and composition, then adapt the grammar to QuickSend rather than copy the original product or assets.
• Opus 5.5 writes the Remotion project, reuses the real product interface, creates missing assets through Higgsfield and renders drafts for review. Jason then uses Motion OS to inspect scenes, change editable properties and turn timestamped notes into one structured prompt for Claude.
• The cinematic workflow deliberately separates character design, location design, the approved first look, the remaining camera angles, animation and final editing. Jason reports that trying to generate everything in one pass produced a visibly weaker result.
• The public Notion guide supplies the exact production order, model settings and creator-reported budget: roughly 1.5 Higgsfield credits for the code-first film and about 860 credits for the advanced film. Pricing, promotions and model availability can change, so estimate a current project before generating.
• The title's comparison with Fable or Astra is Jason's judgment based on his own attempts. The video does not present a controlled benchmark with matched prompts, budgets and review criteria, so this article treats it as a workflow preference rather than a universal ranking.

Credit, publication date and disclosure

The embedded tutorial was created by Jason Lee and published on October 7, 2026. Jason also published the companion Motion Graphics with Opus 5.5 guide and the open-source Motion OS repository used in the tutorial.

Higgsfield sponsored the video, and the guide uses Jason's Higgsfield referral link. The workflow, results and quality opinions are therefore creator-reported. This article preserves that disclosure and separates observable production steps from promotional or comparative claims.

The guide's long prompts are reorganized below as adapted production templates. Use Jason's public Notion page when you need his exact wording, example images and current links. Model names, promotional credits, API prices and connector behavior are time-sensitive and should be checked again before a paid run.

The two-level system at a glance

Level 1 is a code-first product launch film. Claude studies the first thirty seconds of the Rene launch, reads QuickSend's prototype and rebuilds the communication pattern around Qbot, an invoicing assistant. Remotion supplies the timeline, while Higgsfield produces the twelve profile portraits that cannot be created convincingly from interface code alone.

Level 2 adds a cinematic plate. A consistent founder and office are designed as still images, converted into a set of approved camera angles, animated with Seedance 2.5 and combined with the motion system from Level 1. The final film alternates an on-camera introduction, narrated b-roll, large chat bubbles, phone screens and an end-card transformation.

This division matters because code and generative video solve different problems. Code is strong at typography, interface accuracy, repeatable motion and revision. Generative video is useful for people, camera movement and physical atmosphere, but it needs references and shot-level control to stay coherent.

Start with reference grammar, not a blank prompt

Jason recommends browsing launch films on What Ships rather than searching an endless social feed. The Rene film becomes the Level 1 reference because it uses chat bubbles, notifications, portraits, mobile interfaces and kinetic type that map naturally to Qbot.

The Claude Tag launch becomes the Level 2 reference because it combines a founder in an office with motion overlays and product screens. Jason asks Claude to analyze the structure and pacing, then creates a new founder, office, script and product story for QuickSend.

A responsible reference brief names the transferable qualities: duration, information density, rhythm, camera logic, hierarchy and transition pattern. It also states what must remain original: brand identity, copy, product interface, cast, location, music and hero imagery.

Level 1 · Build the launch film in code

Jason's public brief asks for a 45-second QuickSend launch film focused on Qbot. It defines six scenes: a chat cold open, the Qbot introduction, an invoice notification, a client-memory sequence, a payment sequence and a Qbot-to-QuickSend end card.

The brief includes the real prototype URL, the invoice data that must remain consistent, typefaces, brand colors, logo geometry, mascot behavior, camera language and readability adjustments. This is why the result is more specific than a generic request for a polished launch video.

Claude builds the composition in Remotion. The interface, copy and most graphics stay as editable code. Higgsfield's Soul 2.0 supplies the twelve square client portraits. Jason reports about 1.5 Higgsfield credits for those images, with the rest of the film rendered from code.

The first output is treated as a draft. Jason spots placeholders where faces should be, changes the end-card sequence, requests interface sounds and adds licensed background music only after the motion works. This keeps expensive or slow work behind a visible review gate.

Motion OS turns visual notes into a structured revision prompt

Motion OS is Jason's MIT-licensed Claude Code skill and local review player. Claude renders the video plus a reel.json file that describes scenes, elements, copy, timings, fonts, colors and assets. The player runs locally, with no Motion OS account or cloud upload required.

The player can scrub the timeline, expose editable fields, move or resize elements, mark scenes as approved and pin a note to an exact frame and position. Send to Claude compiles all edits and notes into one prompt with timestamps and scene identifiers, which is more precise than a folder of screenshots.

The current repository lists Claude Code, Node.js 18 or newer and FFmpeg as requirements. The global install command is git clone https://github.com/jasonlee-breadcrumb/motion-os ~/.claude/skills/motion-os. For one project only, clone it into that project's .claude/skills/motion-os folder instead.

Review the repository before installation, especially when a skill can run local commands. The included player works with any renderer that outputs a video and reel.json, although the README recommends Remotion. Remotion licensing also depends on organization size, so commercial teams should check the current license rather than assuming it is universally free.

Music and sound are still directed, not magically solved

For Level 1, Jason chooses Be My Lover (Instrumental Version) by I'MIN from Epidemic Sound and saves it as music.mp3 in the project. The brief asks Claude to align the first kick with the Introducing moment, keep the track under the effects and fade it across the end card.

For Level 2, he uses Tickety Toc by The New Fools and identifies the point where the groove begins. Claude is told to land that moment on the founder sitting down, duck the track under speech and raise it for the end card. The motion system also reuses chat sounds, taps, payment dings and mascot effects from Level 1.

These tracks require the appropriate license. Jason's Motion OS repository removes the music from its downloadable example while preserving voice and sound effects. That is the right separation between an open-source project file and a commercially licensed soundtrack.

Level 2 · Plan the cinematic film before generating assets

The advanced film begins with a complete plan, not an image prompt. Claude receives the reference, founder description, office direction, 27-second script, fixed invoice totals, eight-shot list, motion-overlay copy and end card. It returns a shot map that connects each narration line to a specific visual.

Jason then follows a strict order: create the character sheet, create the empty office, combine them into one approved first look, derive the other camera angles, animate every shot and edit the final film. Each approved stage becomes evidence for the next stage.

The order reduces simultaneous uncertainty. When a generation only needs to solve one problem, such as a new angle using an approved person and room, it is less likely to drift than a one-shot request that invents the cast, environment, lighting, performance, camera and dialogue together.

Step 1 · Lock the founder with a character sheet

The guide uses Seedream 5.0 Pro at 16:9 and 2K, with a batch of four. A reference portrait anchors the face, skin tone and hair. The new prompt changes the wardrobe and asks for two panels on a white seamless background: one full-body view and one chest-up portrait.

The prompt specifies an olive-skinned founder around thirty, a sleek low bun, forest-green cardigan, white shirt, navy trousers and brown loafers. It also separates the lens logic: a 50mm full-body view and an 85mm portrait. Neutral light and a blank background make the image useful as a production reference rather than as a finished scene.

Jason reports 2.5 credits per image, or 10 credits for the batch of four. Choose one approved sheet and reuse it throughout the project. Do not let later prompts silently redesign the cast.

Steps 2 and 3 · Lock the office, then approve one first look

The office is generated separately with Soul Cinema at 16:9 and 2K. The scene uses light oak, fluted panels, stocked shelves, a dark grey laptop, a mustard mug, controlled accent colors and soft daylight from camera left. No person appears in this stage.

The first look combines the approved office and character sheet with Seedream 5.0 Pro. It is a wide documentary-interview composition from across the room, with the founder small in frame, the whole desk visible and the light direction preserved.

This hero frame becomes the visual contract for the film. Approve facial identity, wardrobe, body scale, desk geometry, prop placement, exposure and lens feeling before generating another angle. Jason reports about 0.12 credits per Soul Cinema image and five credits for two first-look outputs, but those numbers may change.

Step 4 · Derive the remaining angles from approved references

Claude receives the first look, character sheet and office, then requests over-the-shoulder, top-down, close-up and phone start frames. Each frame uses the approved evidence, which helps the desk, props, lighting and founder remain consistent.

The top-down shot is treated differently because the face is hidden and the camera points straight down. Lens choices are also explicit: 50mm for the over-the-shoulder and phone shots, 85mm for the close-up. The founder is not speaking in these images, so the prompt keeps her mouth closed.

The guide asks for a contact sheet containing every candidate angle. That single review surface is important because continuity problems are easier to see when the entire sequence is visible at once.

Step 5 · Animate in drafts, then finalize selected takes

Each approved start frame becomes a Seedance 2.5 clip in omni-reference mode. The guide enables audio and Draft mode, attaches the start frame plus the character sheet and omits an end frame. Non-speaking clips request restrained physical action and natural production sound.

The founder's complete 27-second voice performance is generated as a separate take. The on-camera shot uses only the opening lines plus a short audio reference from that full take. The remaining scenes use narration extracted from the voice take, which avoids forcing every clip to solve dialogue and physical continuity simultaneously.

Jason reports three credits per second for 480p drafts and ten per second to finalize selected takes to 1080p, with approximately 252 draft credits and 570 finalize credits. Treat these as a snapshot from the guide, not a price promise. Confirm the platform's current model, resolution and finalize rules before spending.

Step 6 · Assemble the cinematic plates and code overlays

The final Remotion edit follows the Claude Tag structure while using QuickSend's original story and Level 1 components. The opening crane sits under a heavy blur and two chat cards. The blur clears as the founder sits, the title appears, then narration and product overlays move across the remaining shots.

Claude cuts the voice take at pauses, aligns the tap with the word send, reuses phone screens, Qbot bubbles, the mascot and sound effects, and finishes on the Qbot-to-QuickSend logo flip. The mix target in Jason's guide is minus 14 LUFS.

Contact sheets remain part of final verification. Review the edit at full size and phone size, check every figure against the script, listen for chair squeaks or abrupt ambience changes and confirm that the final music license covers the publishing destination.

Model and credit cheat sheet from Jason's guide

Soul 2.0 is used for the twelve Level 1 profile portraits. Soul Cinema creates the empty office. Seedream 5.0 Pro creates the character sheet, first look and remaining camera angles. Seedance 2.5 generates the cinematic video clips.

The creator-reported image rates are about 0.12 credits per Soul 2.0 or Soul Cinema image and 2.5 credits per Seedream 5.0 Pro image. The reported video rates are three credits per second for 480p drafts and ten credits per second for finalization.

The guide estimates around 36 image credits plus 822 video credits for the advanced film, or roughly 860 credits in total. Promotions mentioned in the video, including large numbers of free Soul generations, are especially likely to change and should not be used as a durable budget assumption.

What still breaks, and why the workflow works anyway

Jason's one-shot attempt looks less convincing because the model is asked to decide too many variables at once. The staged approach is slower, but it makes continuity inspectable and lets the operator reject one asset without rebuilding the entire film.

The finished direct-to-camera performance still has visible AI mannerisms. Jason says the narration-led shots look stronger than the speaking shot. This is a useful production constraint: use the on-camera moment briefly, then let voiceover, inserts, interface motion and reaction shots carry the rest.

Motion OS improves feedback but does not replace visual judgment. Moving an element in the player may be a live edit or a preview depending on the renderer. Text, timing and structural changes still need Claude, a rerender and a fresh review.

The best result comes from combining strengths: code for exact interfaces and graphics, still-image models for identity and art direction, video models for physical footage, a human for taste and accountability, and a local review tool for precise iteration.

A production checklist you can reuse

Before generation: approve the concept, audience, product claim, reference grammar, original brand assets, script, fixed facts, cast, office and soundtrack plan. Confirm the current model prices and set a maximum spend.

Before animation: approve the character sheet, empty location, first look and contact sheet of all camera angles. Name every file by shot and version so Claude cannot confuse a rejected image with an approved reference.

Before finalization: review the 480p drafts, select exact takes, confirm that finalization preserves those takes and finalize only clips that appear in the edit. Keep a record of credit use and failed generations.

Before publishing: verify product facts, spelling, invoice totals, UI states, visual continuity, lip sync, color, loudness, captions, music rights, creator credits, output dimensions and codec. Save the editable Remotion project, reel.json, prompts, approved references and final manifest beside the master video.