Vi3W
Skip to content

Guides

How to prompt Claude or ChatGPT for a 3D model you'll actually use

Write chat prompts that produce usable 3D models: the brief structure, example prompts for props and characters, and follow-ups that fix results.

6 min read

A pencil sketch of a brass lantern on paper, with the same lantern rising off the page as a solid, lit object
Illustration. A good request reads like a brief, not a wish.

Quick answer

How should I prompt an AI chat for a 3D model?

Treat the request like a brief to a 3D artist: name one object, its purpose, size, style, materials and the detail level you need, then let the assistant restate it before anything is generated. Ask it to show the result and wait, and fix one thing per follow-up message instead of rewriting everything.

Why prompting a chat is different from prompting a generator

When you type into a 3D generator directly, your words go straight to the model that makes the mesh. When you type into Claude or ChatGPT with a 3D tool connected, there's a middle step: the assistant reads your message, decides what to send, and writes the actual generation prompt itself.

That's useful. The assistant can ask questions, catch contradictions, and turn "a cool sword" into a clear description. It's also a risk: it may embellish, add a background, or run three steps you didn't ask for. The way to get good results is to brief the assistant well and keep control of each step.

If you want prompt tips for generators themselves, AI 3D prompts covers that. This guide is about the chat layer.

The brief: six things to say

A good request reads like a short brief to a 3D artist. Cover these, in roughly this order:

  1. The object. One thing. "A treasure chest", not "a pirate's cabin".
  2. What it's for. "For a mobile game", "for a 3D print", "for a product page", "for a VR scene". This decides detail and file format.
  3. Size. "About 60 cm wide." Generators don't know scale, but the assistant can pass it on and you can apply it later.
  4. Style. "Stylised, hand-painted", "realistic", "low-poly, flat colours", "toy-like".
  5. Shape and materials. "Short and wide, rounded lid, dark wood with iron bands and rivets, worn edges."
  6. What happens next. "Show me the result and wait before doing anything else."

Put together:

"Using Vi3W, make a treasure chest for a mobile adventure game. About 60 cm wide. Stylised, hand-painted look. Short and wide with a rounded lid, dark wood planks, iron bands and rivets, slightly worn edges. Show me the result and wait."

A longer template you can save is in 3D model brief for AI generation.

Let the assistant restate the brief

Before the first generation, ask the assistant to say back what it's going to send:

"Before you generate anything, write the exact prompt you'd send to Vi3W and list anything you're unsure about."

This catches problems cheaply. You'll see whether it added "on a wooden table in a dim tavern" (a scene, which generators handle badly) or dropped your size. Fix the text, then approve.

Example prompts that work

A game prop

"Make a stylised wooden barrel with three iron hoops, slightly bulging, mossy at the base, for a cosy farming game. About 90 cm tall. Show me first."

A character to rig

"Make a full-body stylised knight character in a T-pose, arms out, legs apart, chunky proportions, plain armour with a blue tabard. For a game. Show me first, then I'll ask for bones."

A T-pose or A-pose makes Add bones far more reliable, because arms and legs are clearly separate from the body.

A creature

"Make a four-legged stylised fox, standing, tail out behind, orange with a white chest. Show me first. Next I'll want bones for an animal and a run."

A printable figurine

"Make a chunky dragon figurine for 3D printing: sitting pose, wings folded close to the body, thick tail, no thin spikes. Show me first, then convert to STL."

Folded wings and no thin spikes aren't style choices here; thin parts break or fail to print.

From a reference image

"Here's my concept sketch. Describe it as a 3D artist would, then use that description and the image to make the model. Keep the silhouette."

More on images: Claude image to 3D.

A pencil sketch of a brass lantern on paper, with the same lantern rising off the page as a solid, lit object
Illustration. A clear silhouette, in words or in a sketch, is what the generator follows most.

Follow-ups: fix one thing at a time

The first result is rarely the last. The trick is choosing the cheapest fix:

  • The surface is wrong, the shape is right: "Give it a new paint job: lighter oak, rusted iron." This keeps the mesh and changes only the textures.
  • Too heavy for the target: "Run Clean shape at a smaller model size."
  • The shape is wrong: regenerate, but change one thing in the brief. "Same chest, but taller and narrower."
  • It needs to move: "Add bones, then make it walk."
  • Wrong format: "Give me the FBX" or "Convert it to STL."

Avoid "make it better". The assistant will guess, and guesses cost a generation. Name what's wrong.

Words that help, and words that hurt

Helpful:

  • Silhouette words: short, tall, wide, round, chunky, slender, tapered.
  • Material words: oak, iron, brass, stone, leather, cloth, ceramic.
  • Style anchors: stylised, hand-painted, low-poly, realistic, toy-like.
  • Condition: new, worn, chipped, mossy, rusted.

Unhelpful:

  • Camera and lighting: "cinematic lighting", "close-up", "bokeh". A model has no camera.
  • Scenes and backgrounds: "in a forest", "on a table".
  • Several objects: "a sword and a shield". Make two requests.
  • Brand names and characters: they raise copyright problems and often don't help the shape.
  • Vague quality words: "amazing", "ultra-detailed", "8k". Say which details you want.

Rules worth putting in your first message

Paste this at the start of a session and you'll save time and XP:

"Rules for this chat: one object per generation. Before any paid Vi3W tool (generate, Add bones, Make it move, New paint job, Clean shape), tell me what you're about to run and wait for my OK. Show every result before the next step. Keep my size and style unless I change them."

In Claude, you can turn rules like these into a skill so they apply every time; see Claude skills for 3D AI.

What good prompting can't fix

A perfect brief won't make a generator produce exact CAD dimensions, a whole scene, clean hand-made edge loops, or a model of something it can't infer, like the hidden back of an object it's never seen. It won't make a rig that suits every animation either. Prompting improves the odds; it doesn't replace checking the model, and sometimes the faster fix is ten minutes in Blender.

Try a brief now

Write a six-part brief for the next prop on your list and run it through Claude or ChatGPT with Vi3W connected (Claude setup, ChatGPT setup), or paste it straight into the Vi3W app.

Frequently asked questions

text to 3d model prompt

A good text-to-3D prompt names one object and gives its style, shape, materials and use, for example "a stylised brass lantern, chunky proportions, worn edges, for a mobile game". Leave out cameras, lighting and backgrounds; they don't apply to a model.

ai 3d model prompt

Start with the object, then add style, materials and purpose. Be specific about shape ("short and wide", "tall and thin") because generators follow silhouette words closely.

how to make claude make good 3d models

Give Claude a 3D tool to call (a generation connector or Blender MCP), a clear brief, and a rule to show each result before continuing. Then iterate one change at a time.

Last reviewed