Picture this: you feed a single floor plan into an AI, and it hands you back over 100 rendered images, a live website, and a one-minute trailer video. All without you touching a second tool. That’s not a pitch. That’s a real workflow, and I could not stop thinking about it after I saw it.
The person who pulled this off is an architect and AI professional named Tianyu, who shared the whole process on LinkedIn. He ran the entire thing inside one Claude Code session in his Terminal, start to finish. What grabbed me most is how ordinary the starting point was: one floor plan. Everything else, the research, the visuals, the site, the video, grew out of that inside a single conversation with the AI.
This is what people mean when they say agentic workflow. Instead of hopping between a dozen apps, you hand the AI a goal and let it chain the steps together for you. And the creator makes a great point that gets lost in the hype: you probably already have access to this. If you subscribe to Claude (Claude Code or Cowork) or ChatGPT (Codex), the tools are sitting right there.
The step-by-step process he followed
Here’s the exact sequence the author walked through inside Claude Code. I love that each step feeds the next, so nothing gets lost in a handoff.
- Start a session with the floor plan. This is the anchor. Every later decision references it, so the AI keeps context about the actual space instead of guessing.
- Research trending interior design styles. Before generating anything, he had the AI scan what’s popular right now, so the visuals land in a real aesthetic instead of a random one.
- Use the Runway MCP to generate visuals. This connection is what let Claude produce those 100+ images without leaving the session. The MCP acts like a bridge to an outside image tool.
- Instruct Claude to build the website. With images in hand, the AI assembled a working site around them, no separate web builder needed.
- Create a one-minute trailer video. A short promo clip to bring the concept to life and show the space in motion.
- Test and optimize. The final pass, checking that everything works and tightening the rough edges.
Why this matters: the expert never left Claude Code, and every file saved locally on his machine. No copy-pasting between five browser tabs, no lost context, no re-explaining the project to a new tool each time. That single-session flow is the whole point.
What an MCP actually is
The creator leans on something called the Runway MCP, and that term might be new to you. MCP stands for Model Context Protocol. Think of it as a plug that lets Claude talk directly to an outside service, in this case a visual generation tool, without you leaving your session. It’s the piece that turns Claude from a chat window into a hub that can reach out and do things.
That’s the mental shift here. The AI isn’t just answering questions. It’s researching, generating media, writing code, and stitching a project together, all under one roof.
The part that made me sit up
Near the end, the original poster shares a small moment I found really telling. During a live session, a friend asked him, “Tianyu, you don’t use ChatGPT anymore?” He said he froze for a few seconds. He’d been so deep in the work that he hadn’t even noticed the shift.
His honest read: once you get used to the agentic process inside one ecosystem, you reach for other tools less and less. Not because they’re bad, but because the friction of switching disappears when one session can do it all. I think that’s the real story here, more than any single feature.
A few things you can take away today
- Start from your real asset: the floor plan worked because it grounded everything. Whatever your version is, a spreadsheet, a draft, a brand doc, begin the session there.
- Sequence your steps: research first, generate second, build third. Letting the AI gather context before it creates leads to sharper output.
- Try connecting a tool via MCP: this is what unlocks media and actions beyond plain text. It’s the difference between a chatbot and an assistant that ships work.
- Keep it in one place: the value came from not switching. Push how far a single session can carry a project before you split it up.
One more useful tip from the mind behind this: once a workflow proves itself, you can ask Claude to pack it into a reusable Skill and share it with others. So the process you build once becomes something your whole team can run.
And if you’re a ChatGPT subscriber rather than a Claude one, the author’s advice is simple: don’t sleep on Codex. He calls it the best alternative to Claude Cowork or Code, so you can try the same agentic approach with what you already pay for.
I’m genuinely excited about how approachable this makes a workflow that would’ve needed a whole team a couple years ago. If you want to see the full breakdown and the details behind each step, check out the original LinkedIn post. It’s worth the read.