Claude Opus 5.5 Motion Design Studio · Full Course, Prompts and Repositories

Quick Summary

• Movez's central argument is that the prompt is a small part of the result. The studio harness does the rest: deterministic rendering, reference analysis, brand assets, a beat grid, sound, critique and repeatable delivery.
• Opus does not directly output an MP4 in this workflow. It writes a program that can draw an exact frame for any time value. A headless browser captures those frames and FFmpeg encodes the video.
• The simplest route uses one HTML composition with window.seek(t), Playwright and FFmpeg. Remotion is better for React-based series and templates, while HyperFrames offers an agent-oriented HTML and GSAP path.
• One-line showreel prompts are useful as an engine test, but they tend to produce similar results. Product work becomes more distinctive when the brief includes real UI, a reference, a state list, a visual system and explicit banned defaults.
• Motion quality comes from time-addressable springs, deliberate continuity between scenes and a new visual payoff every two to four seconds, not from adding more particles or generic effects.
• Sound is planned on the same timeline as the picture. A supplied track can be measured into beats, downbeats and transient hits, while clicks, pops, thumps and whooshes can be synthesized or sourced and aligned to visible events.
• The strongest projects are iterative. Contact sheets, fast-action strips, phone-size tests, loop checks and repeated scoring make the model inspect and repair its own output before the full export.
• The final step is operational: render 9:16, 1:1 and 16:9 from one timeline, then package the workflow as a skill so future projects begin with a short command instead of another enormous brief.

Credit, publication date and evidence level

The original full course was researched and written by Movez (@0xMovez) and published as an X article on September 27, 2026. Movez also published a readable Substack edition. The framework, source selection and twelve-part structure belong to the creator, and the original article remains the canonical source.

This post is an independent educational reorganization. The prompt pack is newly written from the public methods instead of copying the source article's long templates word for word. Every original X example and repository is linked below, with a smaller set embedded at the relevant step.

The performance figures, production times, model-call counts and client-price comparisons in the examples are creator reports. They demonstrate possibilities, not guaranteed speed, quality, cost or commercial value for another user.

Anthropic officially introduced Claude Opus 5.5 on September 22, 2026. Its release page describes higher effort settings for demanding work, but the motion-design results here come from independent creator experiments rather than an official Anthropic video workflow.

Step 1 · Understand what is actually producing the pixels

The model writes text, usually HTML, JavaScript, React or support scripts. The video is created when that code draws frames and a renderer encodes them. This distinction explains why the surrounding project matters as much as the creative prompt.

The core contract is window.seek(t) or an equivalent draw(t) function. Give it a time value and it must paint the same frame every time. A 15-second film at 60 frames per second needs 900 final frames, plus extra subframes if motion blur is calculated by temporal sampling.

A headless browser loads the composition, seeks to each sample time and captures the canvas or page. FFmpeg receives those frames, encodes H.264 and mixes the soundtrack. Because time is explicit, a revised color, spring or text cue can be rendered again without manually rebuilding a timeline.

Movez notes that Opus often chooses a dependency-light HTML route unless the brief explicitly asks for Remotion, HyperFrames or another framework. Pick the route based on reuse and team needs, not because one library is universally best.

Step 2 · Install a studio, not just a chat window

A useful local stack contains Node.js, FFmpeg, Python for audio analysis and a browser automation layer such as Playwright. The agent needs shell access so it can generate code, render frames, listen to the result, inspect images and repair the project.

Create one project folder per film. A practical structure includes assets, audio, references, documents, prompts, source code and output folders. Keep brand material and approved references separate from generated intermediates.

Put persistent production rules in CLAUDE.md or the equivalent project instruction file. Define deterministic rendering, banned visual defaults, typography limits, color rules, sound targets and the review loop once, then let every future task inherit them.

Use medium effort for narrow corrections and re-renders. Reserve xhigh or max for new visual systems, difficult storyboards and flagship openings. Higher effort can improve planning, but it also uses more time and allowance.

Step 3 · Use the one-line showreel as a renderer test

The viral test prompt asks for a dynamic 15-second motion-design showreel that demonstrates the model's ability and tells it to go all out. Its strength is genre compression: showreel implies fast cuts, varied techniques, a strong opening and best-work-first pacing.

It is a useful smoke test because there is no product fact to misrepresent. If the system can plan, animate, score and export the reel, the basic renderer works. It does not prove that the system can understand a brand, preserve a real interface or communicate a useful product story.

Repeated use creates visual convergence. Many reels default to centered type, gradient backgrounds, corner labels and broad fades. Keep the one-liner for technical validation, then move immediately to references, original assets and a structured brief.

Step 4 · Point the engine at a real brand

A product film needs more than a URL. The agent should collect the real logo, screenshots, fonts, colors, product copy and approved performance metric before animation. List every asset so the operator can catch a wrong or missing source early.

A compact story works well: open with the problem, assemble the product, demonstrate three actions, present one verified proof point and end on the logo plus call to action. Each beat should earn its place and use the real interface rather than a fictional reconstruction.

Keep one session and one source folder per brand. The renderer, sound vocabulary and export scripts can be reused, while the assets, claims and style guide remain isolated from other client work.

Tony Dinh reported creating a TypingMind reel in under thirty minutes after previously paying more than one thousand dollars for similar work. Treat that as his experience, not a universal production estimate or price benchmark.

Step 5 · Name a look and feed the model evidence

A reference can be one frame, a complete video or a folder of your own past work. A still is useful for palette, type, grain and composition. A video adds timing, transition and camera evidence. A personal image library can define a style that competitors cannot copy.

For video analysis, extract frames at a regular interval, then write a style guide and shot grammar before coding. Record exact colors, type roles, shot duration, transition families, camera paths, texture and how text enters or exits.

Tell the model to borrow grammar, not content. Do not reproduce another creator's brand, logo, proprietary footage, characters, music or exact sequence. References should help the system reason about design, not automate plagiarism.

Useful discovery sources include WhatShips for launch videos, Dribbble's motion graphics search and public films from adjacent products. Confirm asset rights separately before using any media in the final piece.

Step 6 · Write the state list, not the vibe

The strongest UI-morph prompts are specifications. They define required inputs, creative direction, an ordered state list, rendering rules, known failure modes and a planning gate before code.

A single container can become a call-to-action, field, loader, confirmation, card, chart, command palette and toast without cutting away. The container changes size, radius and fill while its contents enter and leave on their own timing. A cursor supplies visible cause and effect.

The loop is part of the design. The final container, cursor position and velocity must resolve to the opening state. A visual match without a matching motion state can still create a seam.

The embedded example by zero (@twoclipping) supplied the most widely shared XML-style prompt pattern in the source course. Movez adapts that structure into a reusable UI product-film specification.

Step 7 · Build the deterministic seek(t) engine

Route A is a small browser renderer: one HTML composition, an explicit duration, a canvas or DOM scene graph and window.seek(t). Each scene calculates its state directly from local time. Seeded noise replaces Math.random so rerenders remain stable.

A Playwright render script opens the page at the delivery resolution, waits for fonts, calls seek for every sample and pipes screenshots into FFmpeg. Multiple subframes can be averaged for motion blur, although extra sampling increases render time and can expose ghosting in poor trajectories.

Route B uses a framework. Remotion is a strong fit for React components, data-driven templates and repeatable series. HyperFrames is designed around agent-authored HTML and GSAP compositions. The choice should be explicit in the brief so the model does not silently switch architectures.

Do not put will-change on text or UI that the camera scales if it produces blurry rasterization. Test typography, fonts and media seeking in the actual headless path, not only the live browser preview.

Step 8 · Replace generic easing with time-addressable springs

A fixed ease moves between two values but does not communicate mass. A damped spring accelerates, may overshoot slightly and settles. The closed-form version can be evaluated at any timestamp, which preserves deterministic seeking.

When one property receives several targets, represent every target change as its own timed spring contribution. Do not restart a hidden simulation. This lets the renderer jump directly to frame 812 without replaying the first 811 frames.

Different objects deserve different motion. Buttons and leading edges can be snappy. Cards and cameras can use a moderate response. Large type and logo lockups should feel heavier with little or no overshoot. Mascots can tolerate more playfulness when the brand supports it.

Text inside a morphing container should leave before the next shape transition and enter after the new state has begun. Independent content timing prevents labels from overlapping during fast changes.

Step 9 · Score the motion to a measured beat grid

If the production has a music track, analyze it before storyboarding. Store tempo, beats, downbeats and transient peaks in a JSON file. State changes can land on beats, major reveals on downbeats and interface sound effects on measured visual hits.

If no track is supplied, music and sound effects can be synthesized on the same timebase as the visuals. A small deterministic sound script can create clicks, pops, thumps and whooshes from oscillators and seeded noise, then export a WAV for mixing.

The final mix still needs listening, metering and rights review. A target such as minus 14 LUFS is a delivery starting point, not a substitute for checking dialogue, transient peaks, platform requirements and the exact license for every external track or effect.

Onur Ozcan's Steve Jobs film demonstrates the long-form version: Remotion, React and SVG scenes with a code-generated soundtrack and many transitions aligned to a shared tempo.

Step 10 · Write an overnight director's brief

A flagship brief hires an entire virtual crew. It defines the emotional result, references, tools, budgets, character rules, beat sheet, on-screen text, workflow gates, critique loop and deliverables. It does not ask for a final render before the plan exists.

The production sequence should move from style guide and shot list to stills, contact sheet, low-resolution animatic, full animation, sound and final output. Long projects benefit from one shared animation guide before subagents divide the chapters.

Donald Jewkes reported dictating for about five minutes and letting Claude work for twelve hours on the Claude Pop music video. His public prompt was much longer than the headline one-liner. The useful lesson is structured autonomy with checkpoints, not unattended execution as a universal recommendation.

For hybrid work, a video model can supply difficult character or physical motion, then the code layer can trace, restyle or composite the result. Keep generated assets, permissions, API budgets and provenance explicit in the brief.

Step 11 · Make Opus inspect its own frames

Render a contact sheet at regular intervals, a strip around the fastest action, a phone-width overview and a doubled loop. Each artifact reveals a different failure: weak composition, one-frame pops, unreadable type or a visible seam.

Score the opening hook, phone readability, motion, variety, composition, brand accuracy and sound synchronization. Record the three worst timestamped problems, fix only the affected sequences and generate new evidence. Repeat until the agreed threshold is met.

Mable Joseph's watercolor short is a useful counterweight to one-shot marketing. The creator reported 163 model calls and nearly seven hours of work. Iteration is not proof that the model failed; it is the production method that turns a first pass into a directed result.

Also run a determinism test. Render the same short interval or target frame twice and compare hashes. A mismatch means random state, timers, media seeking or another hidden dependency still affects the output.

Step 12 · Export every format and package the studio as a skill

Build scenes against a layout function instead of hard-coded widescreen coordinates. Render vertical, square and widescreen from the same story logic, but reflow typography and interface composition for each format. Do not treat a center crop as responsive direction.

Package the process as a motion-reel skill. It should collect inputs, gather approved assets, analyze the reference, measure or synthesize music, produce the shot list, build the deterministic renderer, run critique rounds and deliver video, poster and contact sheet.

A repeatable skill can support client work, but do not promise agency-level results, fixed turnaround or revenue based on social examples. Price creative direction, rights, review, revisions and delivery responsibility, not only rendering time.

Chain (@achxvi) publicly demonstrated a longer product film and offered related motion-design work as a service. The example supports the existence of a service opportunity, not a guaranteed customer acquisition model.

Repository map · What to clone and why

PDoomVideo is a complete music-video source with ANIMATION_GUIDE.md, STORYBOARD.md, nine scene chapters, a browser studio and a resumable FFmpeg renderer. It is the best reference for multi-chapter delegation.

ClaudeAnimationBase is the smaller starting point. It packages p5.js, p5.brush, a reusable Clawd character, acted emotions, model sheets, contact-sheet review and a renderer under an MIT license.

buildwithhanif's claude-animation-skill is a hand-drawn Node Canvas route with a detail bible, rigs, brush and watercolor treatment, frame-matched recreation and synthesized sound.

HyperFrames provides the HTML and GSAP framework path. Remotion's official AI skills provide the React and composition path. Choose one when the project needs reusable components instead of a single custom HTML film.

Battle-of-Austerlitz-Film is a long-form example. awesome-ai-motion curates motion work, source code and prompts. athemeroy's collection tracks Claude 5.5 video demonstrations and links each result back to its evidence.

Inspect the current README, license, issues and recent commits before cloning. A public repository is not automatically suitable for every commercial project, dependency policy or security environment.

Verdict · The harness is the studio

The one-line prompt proves that the model can improvise. The twelve-step system proves whether a team can direct, reproduce, review and deliver the result.

The transferable advantage is not one visual style. It is the combination of an explicit render contract, real brand evidence, timed states, sound, frame-level critique and packaging. Once those pieces are stable, the next project can start from a short command without losing the discipline of the full course.