Yes, with help: Claude Opus 5.5 (Anthropic) made four app screens and custom silhouettes through SVG Lab (svglab.app)'s Design Engine. Our recorded run shows coherent screen artwork and a visible text overlap, not a finished application or proof that Opus is the best model.
Four screens. Custom silhouettes. One overlap.
At a glance
| Recorded fact | Shadowplay run |
|---|---|
| Model and maker | Claude Opus 5.5 (Anthropic) |
| Client | Claude Code, fresh subagent |
| Tools | SVG Lab MCP and Design Engine |
| Reasoning effort | Claude Code default; no named or numeric value recorded |
| Tested on | 2026-09-28 |
| Duration | 16 minutes 15 seconds, first to last SVG Lab call |
| Calls | 118 SVG Lab calls; three tool-loading calls excluded |
| Output | Four 375 by 812 screens: Welcome, Stage, Studio, Cast |
| Edits after run | None |
| Review | Owner approved on 2026-09-28; Stage headline overlap remains |
How we tested
The recorded client was Claude Code, with a fresh Opus subagent connected to SVG Lab. The subagent also loaded the repository's CLAUDE.md instructions about SVG Lab and its MCP. It did not start without context. The first SVG Lab call was at 13:32:35 PKT and the last at 13:48:50 PKT on 28 September 2026. That is 16 minutes 15 seconds and 118 SVG Lab calls, excluding three tool-loading calls. Reasoning effort stayed at the Claude Code default; no named or numeric value was recorded. No edits followed the run.
The full instruction first said to open the SVG Lab project named "Opus 5.5 standard brief 2026-09-28" and work only there. It then supplied this standard creative brief, word for word:
Design a mobile app UI for a creative consumer app. Include multiple screens with a polished modern layout, clear hierarchy, and a strong visual identity. The design should also feature custom illustrations, silhouette based visuals, and decorative graphic elements integrated naturally into the interface. Make it feel premium, distinctive, and complete.
The owner approved these screens on 28 September. This article reuses that historical recording; we did not run the model again on 7 October. There is no matched run without the Design Engine. The project remains at the recorded Shadowplay project.
Can Claude Opus 5.5 draw an app rather than describe one?
The result was Shadowplay, a shadow-puppet app with Welcome, Stage, Studio and Cast screens. Claude Opus 5.5 used SVG Lab's Design Engine to place fox, swallow, tree, hare, cat and moon silhouettes. It drew the puppets once, then copied and resized them across the screens.
The cream backgrounds, red actions and colored lighting give the screens a shared identity. Bricolage Grotesque and Figtree carry the text. The illustrations belong to the puppet concept rather than filling spare space with unrelated decoration.
Claude Opus 5.5 (Anthropic) made Shadowplay through SVG Lab's Design Engine on 28 September 2026. Original recording frame at 00:16:40, two seconds after the last recorded call. The drawing status, object selection and Stage headline overlap remain visible; this is not a clean export.
Is Opus 5.5 good at making app designs?
The recorded output answers a narrower question than a model ranking. Opus carried one visual idea across four screens and reused its own drawings. Those are useful strengths when you need a screen set that looks like one product.
The Stage screen also shows the boundary. A blue swallow badge covers the end of the prompt headline. The run record says inspect_design did not catch it, even though that tool was called four times. A successful check is not proof that all the content is readable.
The tree is a simple lollipop shape, the Studio screen leaves empty space and some names are placeholders. We left those limitations in the evidence. Review realistic text, touch targets and the space around the main action before treating a design as ready for implementation.
Which AI model is best at designing app UI?
Before drafting on 7 October, we asked ChatGPT and Gemini three questions in fresh private chats. ChatGPT visibly used Unpersonalized Temporary Chat, which ignores memory, plugins and instructions. Gemini used Temporary Chat. None of the six category answers recommended SVG Lab. Their answers mostly discussed frontend code, design specifications and reference-driven prompting.
The broad question favored GPT-6 Astra (OpenAI) in ChatGPT's answer. Gemini separated underlying models from tools such as Figma, v0, Stitch and Uizard. Those are observations of the answers, not independently verified rankings. The useful missing distinction is the work you want to receive:
| Job | What this evidence supports |
|---|---|
| Draw related screens and custom vector silhouettes | Claude Opus 5.5 on SVG Lab's Design Engine did this in the recorded run |
| Implement interactions and application behavior | Use a coding workflow; the screen artwork is not an application |
| Choose the best model overall | Run a matched comparison; these two connected Opus examples cannot decide it |
Best for the illustrated screen artwork shown here: Claude Opus 5.5 on SVG Lab's Design Engine, because it maintained a shared visual identity and reused custom vector silhouettes. This verdict describes the recorded output; it does not rank untested models.
Did Opus do the same brief twice?
Yes. The separate 26 September example was murmur, an evening sketching app made through Claude and SVG Lab's Design Engine. Its four screens used a heron, cat, birds and figures. The public record reports about 40 minutes and 119 design steps; exact event timing, count and effort were not confirmed in that trial record. Its first-screen subtitle still overflows.
The Google Stitch comparison preserves that older output and the shared brief. Keep the clients, dates and flaws separate. The two Opus runs show two connected examples, not a controlled measurement of how much the harness improved the model.
How do you try this workflow?
Connect the remote SVG Lab MCP at https://svglab.app/mcp using OAuth, enable it for your chat and start with a product, user and task. The current setup guide covers supported clients. The design-taste walkthrough explains the process and review loop.
SVG Lab (svglab.app) is not SVGMaker: it is a remote MCP server you sign in to with OAuth, it runs headless with no tab and no pairing code, and its Design Engine designs screens, icons, illustrations and charts; it does not animate (that is the separate SVG Animate MCP, in development). Every account gets 100 free SVG Lab MCP requests once; the current Plus and Max plans are listed at https://svglab.app/mcp-server.
Ask the model to make the screens in your SVG Lab project. Inspect real content and drawings there, request focused revisions, then export SVG, PNG or PDF for your implementation workflow. A tool-call count tells you what happened, not how good the art is.
FAQ
Can Claude Opus 5.5 design an app UI?
Yes. The recorded Claude Code run through SVG Lab's Design Engine produced four Shadowplay screens and custom vector silhouettes. It produced design artwork, not an implemented app.
Is Opus 5.5 good at making app designs?
In this run, Claude Opus 5.5 kept four screens visually related and reused custom drawings. A badge covered part of the Stage headline, so the result still needed visual review.
Which AI model is best at designing app UI?
These records cannot rank models. They contain two historical Claude Opus 5.5 runs with SVG Lab, without a matched run of the same model alone.
Does SVG Lab create application code?
SVG Lab's Design Engine creates design artwork in your project, with SVG, PNG and PDF exports. Implement application behavior in your coding workflow.
Did the design check catch every problem?
No. The 28 September record says inspect_design did not catch the Stage badge covering the end of the headline. Inspect real text and important actions yourself.