Building 3D in Blender from a written brief to an AI

Keel, ribs, planking, deck, cannons, masts: a ship assembling piece by piece, requested from an AI in words. @Ayu_AI_0912 shared the full prompt they used to have Opus 5.5 and Blender make drawings and a video of a pirate ship ¹.

Watch the finished video in the post first.

The creator writes (in Japanese) "Tried making it with Opus 5.5 x Blender" ². The interesting part here is the request, not the result. When you ask an AI for 3D drawings and a video, what exactly do you write, and how far do you spell it out? Let's walk through the design of the prompt.

The drawing and the video read the same dimensions

The request starts with a fictional pirate ship: make CAD drawings and a Blender video ³. The prompt asks for 2D DXF only, in mm, with an A2 drawing frame and a title block ⁴. It also asks for a note stating that this is an image of a fictional ship and not a drawing of a real one ⁵.

The key to the design is how the dimensions are held. The prompt specifies that the dimensions are decided in a single Python shape module, and that both the drawings and the video load it ⁶.

If the person drawing and the person animating each keep their own dimensions, they can drift apart. The prompt is built so the source data lives in one place and both the drawings and the video read it ⁶. It is the old habit of keeping a single source of truth, written into the request from the start. That is the most interesting part of this example.

Even the order of assembly is written down

The video spec is just as precise: Blender's EEVEE, 1920x1080, 24fps, a 20-second main section plus a 3.5-second logo, as an MP4 with sound ⁷.

The order of scenes is specified too. According to the prompt, the ship goes together as keel, ribs, planking, deck, cannons, masts, yards, rigging, then sails ¹. The next scene is specified as well: the paper vanishes, the sea fills in from below, the ship floats, and the camera pulls back to a full view under a sunset sky ⁸.

Nowhere does it just say "make a cool pirate ship video." It puts into words what to show, in what order, and how, including camera moves. The more you hand to an AI, the more your job shifts from drawing pictures to designing procedures.

Sound, synthesized in Python with no outside assets

Another standout is the audio. The prompt specifies that sound is synthesized in Python with no external material: a sailors' song, a mallet, sails, waves, seagulls, a ship's bell, and cannons ⁹. Instead of hunting for assets to paste in, the sound is made in code too, so the whole job runs as a single flow with the drawings and the video.

Check first, then ask permission

The "how to proceed" section is also worth reading.

• The prompt tells the AI to look at the drawing preview images and fix overlapping text [[C192]]
• It tells the AI to render a few test stills and check them before the real render [[C193]]
• It tells the AI to run the final render with light settings, letting the PC rest between runs, and to state the expected time and get approval before starting [[C194]]

Instead of fixing things after the work is done, the drawing is checked at the drawing stage and the video at the still-image stage. Before the final render starts, the AI has to say how long it will take and get a go-ahead ¹⁰. A good screen helps when you want to study test stills and drawing previews closely.

A game of never touching it yourself

Others are trying something in the same direction. @Yokohara_h posted a work-in-progress test of building a Blender scene with AI alone, without ever touching Blender personally ¹¹.

The creator writes (in Japanese) that the rule is AI only, modeling included ¹², and that they combined methods to get as much quality as possible ¹³. The post does not say which AI was used.

@KOSAKA_Jun posts a single sentence: a 3D movie generated automatically by AI alone, with Opus 5.5 and BlenderMCP ¹⁴. The post is short and gives no steps or environment.

For finishing touches, see another article

For what comes after the request, there is a tutorial by @KanaWorks_AI. The creator writes (in Japanese) that handing Blender over to Opus 5.5 wholesale usually stops at "plausible" ¹⁵, and lists, among what helped, showing the AI its own render and having it measure with numbers ¹⁶.

Another tip is to have a different AI, not the one that made it, do the scoring ¹⁷. For more, read our article on handing Blender to an AI agent, built around the same post.

One example of connecting an AI to Blender

One example of a bridge between an AI and Blender is MCP for Blender (formerly blender-mcp). Its GitHub README describes it as connecting Blender to any LLM ¹⁸. Listed features include the AI writing Blender Python directly ¹⁹ and using renders so the AI checks its own work ²⁰. It is a third-party integration, not made by Blender ²¹, under the MIT license, and its PyPI package has been renamed mcp-for-blender ^{22 23}.

@KOSAKA_Jun's post says "BlenderMCP" ¹⁴ but does not say which implementation, so it cannot be equated with the MCP for Blender above. The posts by @Ayu_AI_0912 and @Yokohara_h do not say what connected Opus and Blender.

The brain is in the cloud, Blender is at hand

Here is the split. For Claude Opus 5.5, the official availability is the Claude Platform and cloud services such as Amazon Web Services ²⁴. What you can run on your own machine is the Blender side ²⁵: Blender is released under the GPL ²⁶ and is compatible with Windows 11 and others ²⁷.

The posts do not say which PC the rendering ran on. All we know is that the prompt says to let the PC rest between runs ¹⁰. Waiting on the final render and the load on the PC both come back to how much headroom your own machine has.

If you try this at your own desk

NVIDIA says RTX Spark's up to 128GB of unified memory handles 3D scenes up to 90GB ²⁸. The same page says RTX-exclusive features such as OptiX and DLSS keep the ray-traced viewport moving ²⁹.

On the Blender side, the official download page says Blender is compatible with Windows on Arm ³⁰. NVIDIA announced that Blender is adding DLSS 4.5 Ray Reconstruction, arriving this fall with RTX Spark ³¹. The Japanese NVIDIA blog says it is planned for Blender 5.3 ³², so this is a plan, not something shipping. None of this means the tools in these examples run as-is on an RTX Spark machine, so check the supported environment of whatever you want to run in the official docs before choosing.

Here are examples of RTX Spark machines.

If you spend a lot of time spinning the scene around to check its shape, an input device like a 3D mouse is worth a look.

Read next

For another way an AI builds a 3D world, the playable-worlds example reads as a continuation of the same craft of asking. For Blender finishing tips, head back to our Blender agent article.