# A Step-by-Step Workflow for Agentic 3D Modeling

In this tutorial, I treat GPT-6 Astra as a next-generation multimodal agent described by the premise: it can interpret 3D design intent, generate modeling scripts, call professional software and AI generators, and revise the result over several loops. Shapezo is a focused map-to-model tool: I select an area on a map, and its AI generates an initial 3D context model.

The goal is not to automate final delivery. The goal is to create a workflow that is quick to inspect and honest about uncertainty.

## **Step 1: Write the design question**

I start with one narrow question. Examples include: “How should this building step with the slope?” or “Can the public route remain visible from the transit street?” A question about object geometry points toward Astra. A question about geographic relationships points toward Shapezo. If I have both, I keep two linked tasks.

I also define the expected output: a massing script, a contextual scene, a neutral render, or a comparison view. This prevents the agent from generating detail that the decision does not need.

## **Step 2: Prepare the Astra brief**

I give Astra the reference material it needs: sketches, photos, plans, a rough model, known dimensions, and written constraints. I ask it to summarize the intent before writing code. The summary should name the footprint, height logic, openings, circulation, materials, and any assumptions inferred from the images.

![](https://cdn.hashnode.com/uploads/covers/6a912038d49e2dc741ccf326/f0dd218d-6d8e-452d-a033-546e1443504d.png align="center")

![](file:///C:\Users\admin\AppData\Local\Temp\ksohtml3248\wps158.png align="center")

Next, I ask for a small script with named parameters. Useful parameters might include footprint width, floor height, roof pitch, window spacing, and courtyard ratio. I require the script to save a versioned scene and to report warnings when an input is missing.

## **Step 3: Run in an isolated modeling scene**

I run the generated script in a temporary Blender or equivalent scene. I check object names, origin, units, polygon density, normals, UVs, and material slots. I change one parameter at a time and confirm that the geometry responds as expected. If a parameter exists but does not affect the intended object, I treat it as broken until fixed.

I keep the original script and the executed scene together. A rendered image alone is not enough to reproduce the result.

## **Step 4: Create the Shapezo context**

I open Shapezo and draw a boundary around the project parcel plus the surrounding streets and blocks that influence it. I generate the initial AI model and save the boundary, map date, orientation, and three views: aerial, oblique, and eye-level.

![](https://cdn.hashnode.com/uploads/covers/6a912038d49e2dc741ccf326/3ffd56e9-1a35-4704-bc9a-5d3dd41ea16e.png align="center")

![](file:///C:\Users\admin\AppData\Local\Temp\ksohtml3248\wps159.png align="center")

I validate the obvious context: road continuity, relative building heights, open-space connections, major terrain changes, and visible barriers such as rail or water. I create an assumptions list for simplified roads, estimated heights, missing utilities, vegetation, and uncertain property edges.

## **Step 5: Align and combine**

I export a simplified Astra asset and place it into the Shapezo context, or recreate the Astra massing as a neutral block. Before comparing options, I align units, origin, ground level, and north direction. I save the aligned copy separately from the source scene so the transform remains traceable.

I use fixed cameras for every option. From those views I inspect frontage, apparent height, pedestrian approach, service access, shadows, and the relationship to parks, water, or transit. I do not let a new camera hide a weak edge condition.

## **Step 6: Add provenance and review gates**

I label each major element as observed, imported, generated, estimated, manually edited, or rebuilt. I store the prompt, script version, map boundary, source dates, scale corrections, and manual changes. GPT-6 Astra may summarize the state, but I still verify the summary against the files.

Before moving forward, I ask three questions. Can another person reproduce the scene? Can I identify which geometry is approximate? Can I explain what decision the view supports? If any answer is no, I improve the record or rebuild the relevant object.

I also check that the document title, file names, and revision date agree. Small metadata errors can make a correct model look like the wrong version when it is passed between tools or reviewed by a different team.

## **Step 7: Move selected work into production tools**

When the concept survives review, I rebuild geometry that carries real consequences in Blender, Revit, Civil 3D, or the project-standard environment. The Astra script and Shapezo scene remain references for intent and context. They do not become automatic proof of structural accuracy, topology quality, engineering compliance, or final acceptance.

## **Final rule**

Use Astra for multimodal direction, scripting, and cross-tool iteration. Use Shapezo for a fast geographic frame. Keep state, provenance, and human review between every major step. That is how I would gain the speed of an agentic workflow without confusing a plausible 3D draft with a finished design model.