Fable 5.2 turns 2D floor plans into walkable 3D tours: the archviz workflow taking off on X

What was shown. On Nov 10 Chris (@everestchris6) posted a 10-second clip of a flat blueprint turning into a lit, furnished 3D interior, made with Claude Fable 5.2, along with the full prompt he uses. It passed 6,800 likes, 8,000 bookmarks and about 680,000 views. On Nov 11 @RoundtableSpace re-shared it as 'one-shot a tool that builds an entire 3D walkthrough from a 2D blueprint'. His pitch is aimed at people who sell spaces from drawings: home builders and developers selling off plan, architects presenting to clients, interior designers and remodelers. He says these walkthroughs sell for $5k to $10k per plan. That's his figure, not a market rate we've checked.

How the build works, per his prompt. It's a good template for any code-built 3D piece, not just floor plans. (1) Trace first: the agent turns the drawing into one data file of walls, openings and rooms, draws it back over the original and measures the overlap before any 3D work, so scale errors get caught early. (2) Build only the rooms that sell (entry, main living space, kitchen, patio, yard) and close the doors on the rest. (3) Everything is code: Blender via Python for the shell, furniture and textures, three.js for the viewer, and no downloaded models or textures. (4) Bake the lighting with a film-quality renderer so the browser only has to display it, with direct sun in every interior chosen by measurement rather than by eye. (5) 'Fidelity isn't appeal': style each room so it looks photographed. The client's listed upgrades become options you can step into, with a running price and a lead form.

The critic loop is the reusable trick. The builder never grades its own work. After each room, renders plus the client's photos go to a fresh critic that has seen none of the work. It returns yes or no, at most five ranked problems each with a numeric fix, and what's already good. The builder fixes the worst problem, renders again and sends the result to a new critic, repeating until it gets a yes. For matters of taste, a critic picks between labelled A and B versions. This is the same 'agents can't watch video, but they can judge stills' idea behind contact-sheet review in motion skills, made stricter by using a new critic every round.

Where video comes in. Before any 3D work, the prompt has the agent make a cheap sample of the prospect's own plan: three rooms rendered with GPT Image 2.5 matched to their photos, or a short walkthrough clip from Seedance. The paid deliverable is a hosted, phone-friendly page, but the same Blender scene renders straight to fly-through video for listings, socials or a pitch reel. That's the part most relevant to video teams and agencies. The prompt also has the agent find leads by scraping public business listings and draft outreach emails. Those drafts go to you for approval before anything is sent, and the prompt says to use only public pages and never work around bot checks. Whatever tools you use, follow their terms and local outreach rules.

Caveats. This is one creator's demo and prompt, and a 10-second clip doesn't show how the full tour holds up. Expect plan-tracing errors on messy scans, long Blender bake times, and Fable 5.2 costs that add up on long agent runs (Anthropic's own model guide suggests saving Fable for the hardest steps). We haven't reproduced it.

Signals (HOT by our rule): the author's own post took off (about 6.8k likes and 680k views since Nov 10) and a tracked curator picked it up (@RoundtableSpace, Nov 11), within 72 hours.

Our take: for a video shop, the realistic first step is one room, not a whole business. Take a floor plan you're allowed to use, have Claude Code trace it to a data file, build a single room in Blender with Python, run the fresh-critic loop against a reference photo, and render a 10-second dolly move. If that holds up, it's a sellable archviz deliverable that doesn't depend on generator credits.