A custom 3D website is not just an AI coding task. In one agency-site build, Claude Code helped produce interactive pages while Higgsfield turned existing client photos into widescreen motion. The choices that gave those outputs a purpose—who the site should reach, what its visuals should mean, and whether it worked beyond a desktop screen—remained human decisions.
Set the brief before generating the interface
The project had three intended audiences: business owners considering an AI transformation service, potential sponsors, and AI agency owners or consultants. Those groups gave the coding agent more useful direction than a request for a striking homepage alone. The service positioned itself as compressing seven months of enterprise AI transformation into seven calendar days, so the site also had to explain a specific offer rather than merely display effects.
A practical starting brief for Claude Code includes the audience personas, value proposition, and brand voice. The next decision is conceptual: what should visitors see that connects the visual design to the business? Here, the answer was a pulsating 3D orb representing separate business data sources joining an intelligent core. A retro-style opening page dissolved into the darker, interactive experience. The orb responded to cursor movement and sat among floating nodes associated with workplace tools.
That concept gave the coding work a direction, but it also introduced a trade-off. An elaborate opening transition may impress a visitor or delay someone who wants an immediate explanation. Human review has to decide whether the effect serves the audience; generating the effect does not settle that question.
Make the media specific to the work
The build used photos from client engagements rather than default stock imagery. Some projects lacked landscape footage suitable for wide case-study cards. Through a Model Context Protocol (MCP) connection—a way for Claude Code to call an external tool—Higgsfield expanded existing 9:16 vertical photos to 16:9 widescreen images. This outpainting adds imagery beyond a photo’s original edges. The expanded stills then became inputs for short video loops.
▲ Vertical-to-widescreen image expansion
Model choice needed another review pass. The same source images were tested with Cinema Studio 3.0, Kling 3.0 Pro, MiniMax HQ, Seedance 2.5, and Veo 3.1 Ultra. Cinema Studio 3.0 was selected for its color grading and cleaner movement in this build. Other results could introduce too much motion or distort the source image. The useful method is to compare outputs against the original material, not assume that the first generated clip is ready to publish.
The same standard applies to words. Generic cards, stock visuals, and unedited AI copy can make an otherwise elaborate site feel interchangeable. Claude Code can draft and implement quickly, but the site’s claims and case-study descriptions still need writing or substantial editing for the intended buyers.
Turn references into an interface, then test it
For interface direction, the build drew on Mobbin for website patterns and 21st.dev for components. Both were accessible from Claude Code through MCP connections, allowing searches for layouts and component code within the coding workflow. Multiple sphere designs, card layouts, and navigation treatments were generated before selections were made. Three.js, a JavaScript library for rendering interactive 3D graphics, powered the orb’s real-time movement and camera behavior.
Those pieces became more than a hero graphic. The site paired the orb with client-showcase cards that opened into media and project summaries. A separate methodology section explained five parts of the service: auditing, data centralization, workspace building, custom app development, and AI agents. Iterative feedback refined the animations and transitions across these sections.
This is where the division of labor becomes clear. AI tools can produce variations and implement a chosen interaction. A person still has to judge whether the navigation is understandable, whether a case study says enough, and whether motion competes with the offer. Desktop approval was not the end of the build: complex 3D effects needed a dedicated mobile-responsive pass because phone layouts and performance posed different problems.
Treat publishing and measurement as part of the build
A visually ambitious site can lose conventional form conversions if effects obscure the next action. This project added Ospry, a visitor-identification platform, to measure interest beyond form submissions. Ospry claims it can identify approximately 40% of anonymous US visitors; the stated estimate for accurate, verifiable matches among resolved profiles is 80% to 85%. Those figures describe the platform’s claims and estimates, not a guaranteed outcome for every site or region.
Claude Code used Firecrawl, a tool that retrieves web pages as usable text, to read Ospry’s integration documentation. It then updated the site repository with the required tracking script. A dashboard check confirmed that the published site was receiving visitor events. That verification matters: generated integration code is only useful when it works on the deployed site.
▲ Visitor identification and fit scoring
After traffic reached the site and mobile refinements were completed, the dashboard identified 194 visiting companies within 24 hours. It also displayed company information, session activity, and fit scores. An Ideal Customer Profile, or ICP, score rates how closely a company matches chosen customer criteria; the build used a score of 70 out of 100 as a suggested threshold before automated outreach. Ospry’s resolved data could also be routed to a customer database or a tool such as Expandi for follow-up.
Visitor identification raises a separate judgment call. Its effectiveness is described as stronger for US traffic, and its use requires due diligence on applicable privacy requirements. Filtering by relevant company attributes and roles can also keep outreach focused. Neither an identified visit nor a high fit score proves that someone wants to buy.
What to carry into the next build
Start with the audiences and the offer, then choose a visual concept that helps explain them. Use Claude Code to explore and implement options, and Higgsfield to adapt relevant source material when the available formats fall short. Compare generated media, edit the copy, test the interaction on phones, and verify any measurement tools after deployment. AI can shorten production and revision cycles; the standard for a useful website still depends on human decisions at each checkpoint.