Claude Opus 5.5 · Precise Motion Graphics for AI Video

Quick Summary

• Tao Prompts uses an AI video model for cinematic subjects, environments and movement, then uses Claude Opus 5.5 for the elements that need exact control: text, routes, metrics, annotations, HUD displays and timing.
• The workflow begins with clean footage. Do not ask the video generator to render dense interface graphics or important copy into the shot. Leave enough negative space and stable regions for the overlay pass.
• Artlist MCP connects the creative platform to Claude. In the demonstration, Claude generates reference images, reruns them with GPT Image 2.5, animates the selected frames with Seedance 2.5 at 1080p and keeps the assets organized inside one conversation.
• Opus 5.5 then analyzes each clip and writes the motion-graphics layer. This preserves readable text and lets the creator control what appears, where it sits, when it enters and how multiple elements move together.
• The use cases include a city-to-city travel route, dense science-fiction HUD displays and a financial explainer that adds labels to an otherwise plain animation.
• Claude can add sound effects after the visual pass. Tao Prompts reports that the full HUD treatment took about 30 minutes in his test, but render time will vary with clip length, hardware, model load and the complexity of the requested code.
• The method still has limitations. The creator notices small frame-to-frame wobble in some coded animation, and every label, number, route, sound cue and edge alignment needs a final playback review.

Credit and the central idea

This article is based on Tao Prompts' embedded video, published on September 27, 2026, and the supplied transcript. The demonstration asks a useful production question: which parts of an AI video should remain generative, and which parts need deterministic control?

AI video is strong at atmosphere, camera movement, character motion and visual effects. It is less dependable when a shot must contain exact words, a correct route, multiple synchronized interface panels or several values that stay readable. Tao Prompts assigns those jobs to Claude Opus 5.5 as a second post-production layer.

The result is not one model replacing the other. It is a division of labor. The video generator creates the plate. Claude analyzes that plate, writes the overlay system and composites information on top of it.

Why a two-layer workflow is more controllable

Tao Prompts first tries to generate a travel map entirely inside an AI video model. The broad concept appears, but the route begins in the wrong country and the display text becomes unreadable. In another test, the generated phrases differ from the exact wording in the prompt.

These are creator tests, not a benchmark of every video model. They still expose a common production risk: a visually attractive shot can fail if the information is wrong. When the label, number or location carries meaning, approximate rendering is not enough.

Separating the layers creates a cleaner contract. The base clip only needs believable movement and composition. The overlay pass owns spelling, hierarchy, timing, position and information design. If the copy changes, the creator edits the graphics layer instead of regenerating the entire cinematic shot.

Three use cases shown in the video

The travel example traces a route from city to city. Each stop triggers a holographic landmark and a metric card. The important controls are the path order, city name, landmark, number and exact moment of each reveal.

The cinematic HUD example places several panels around a character and later treats a close-up like a helmet interface. The graphics identify the setting, year and robotic subjects while the original footage supplies the street, face and camera movement.

The financial explainer starts with a clean animation about cash, digital payments and international networks. Claude adds readable annotations and effects on top. This pattern also fits product demos, technical explainers, documentary maps, sports analysis and any sequence where visuals must carry verified information.

Connect Artlist MCP to Claude

Artlist MCP lets Claude generate images, videos, voiceovers and music through the user's Artlist account. The official setup is simple: copy https://mcp.artlist.io/mcp, open Claude Settings, choose Connectors, add a custom connector named Artlist, then sign in and authorize the account.

Artlist says the connector can handle text-to-image, image-to-image, text-to-video, image-to-video, voice generation and generation-status checks. Every output uses the same plan permissions and credit pool as the Artlist Toolkit. Unlimited-generation allowances do not apply through MCP, and a started generation cannot be canceled for a credit refund.

The connector is the asset pipeline in this workflow. Claude keeps the brief, references, model choices and generation history in one conversation, while Artlist runs the selected image or video model and saves the output to the associated account.

Step 1: generate images and clean video plates

Tao Prompts uploads a portrait reference and asks Claude to plan three shots. The first image pass uses Nano Banana Pro, but he dislikes the realism and requests a new pass with GPT Image 2.5. This is a useful checkpoint: choose the still frames before paying to animate them.

He then asks for Seedance 2.5 at 1080p. The motion brief deliberately avoids music and leaves room for graphics. General environmental sound is allowed, but the interface sounds are postponed until the final overlay timing is known.

A good clean plate has a clear subject, predictable motion, stable lighting and open composition. Avoid baking labels, charts or interface panels into the generation prompt. Reserve screen space for the later graphic, and keep important faces or objects away from the area where text must live.

Step 2: analyze the footage and build coded overlays

Once the three clips are ready, Tao Prompts asks Claude to inspect them and propose graphics. The walking shot receives location, setting and year information. The face close-up receives an Iron Man-inspired helmet display. The robot scene receives detection markers, information panels and targeting-style graphics.

Claude's advantage here is code. Text strings, colors, sizes, coordinates, easing, opacity and entry times can be expressed as editable values rather than painted into generated pixels. The system can animate several elements while keeping the words legible and the structure consistent.

The prompt should name the communication goal before describing style. Start with what the viewer must learn, list the approved text and values, identify the object or region each graphic tracks, then define the visual language. Style without an information hierarchy produces impressive clutter.

Step 3: add sound after the graphics are locked

The first composite has the requested overlays but lacks interface sound. Tao Prompts sends a short follow-up asking Claude to add effects to the motion graphics. Because the entry and exit timing already exists, the sound cues can follow real visual events instead of guessing where an animation might land.

Use small sounds as feedback: a soft activation click, scan pulse, data sweep, lock-on cue or low transition hit. Keep them below dialogue and avoid assigning a sound to every moving object. One cue can support a group of related elements more cleanly than a stack of competing effects.

If the base clip already contains useful ambience, preserve it and build around it. The transcript says no music was requested during generation, which leaves space for the creator to choose music after the edit and control the final mix.

The final quality-control pass

Readable text is only the first check. Pause on every important frame and confirm spelling, numbers, route order, units, labels and safe margins. Watch the full clip for tracking drift, overlays that slide away from their subject, elements that cross the face and transitions that arrive before the narration introduces them.

Tao Prompts points out a small wobble in some frame-to-frame animation. A coded overlay can still inherit unstable tracking or timing from the underlying analysis. Reduce that problem with fewer anchor points, smoother interpolation, short hold periods and a deliberate tolerance for motion rather than snapping to every frame.

Also verify the composite at delivery size. Dense HUD work that looks clear on a desktop may collapse on a phone. Remove decorative microtext, increase contrast, limit simultaneous panels and keep the important phrase readable in a one-second glance.

A practical production playbook

1. Write the information the viewer must understand. 2. Split that information from the cinematic image prompt. 3. Generate still candidates. 4. Approve the composition. 5. Animate clean plates with negative space. 6. Review motion before building graphics. 7. Give Claude the exact text, data and tracking targets. 8. Composite the overlays. 9. Lock visual timing. 10. Add sound. 11. Run frame-by-frame and phone-size QA.

Keep each layer recoverable. Store the source images, raw clips, approved copy, overlay code, sound effects and final composite separately. If one label changes, the team should be able to rerender the graphic without regenerating the expensive base footage.

The three prompts below translate the demonstrated method into a reusable brief. They are editorial adaptations, not Tao Prompts' private prompt text.