AI for CAD Tools
Why regenerating a CAD part from a prompt changes small details, and how carrying the part into a new conversation keeps the geometry you already approved.
·
⏱
7 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
Regenerating a part from a prompt rebuilds it from a description, so small details such as fillets, counterbore depths and slot positions can shift between sessions. Carrying the part forward starts the next piece of work from the exact geometry you already have. Leo AI added this for parts created in a different conversation, shipped in the week of September 29 to October 4, 2026. Test any tool by measuring three details before and after a transfer and after one edit, and treat a carried part as a working copy until someone has reviewed it.
You asked an AI tool for a bracket, and after a few rounds of corrections it was right: the right hole spacing, the right fillet, a slot that cleared the fastener. Then you opened a new session to build the next piece around it, and the bracket had to be described all over again. The second version came out close, but not the same. The fillet was a little different, the slot a little narrower, and now you are comparing two parts that were meant to be one.
This is one of the quieter costs of generating geometry from prompts. A part is only useful if it stays the same part. This post explains why re-prompting drifts, what carrying a part forward changes, and how to test whether a tool really holds small details in place.
Why the same prompt gives you a different part
A text prompt is a description, not a model. The tool reads it, makes a series of choices where the description is silent, and produces geometry. Run the same description twice and the choices can land differently, because nothing forces the second run to repeat the first. If you want to see where those choices are made, our explanation of how text-to-CAD works from prompt to parametric model walks through the steps.
Most of the drift lives in the details that engineers care about most. A 3 mm fillet can become 2.5 mm. A counterbore that was 6 mm deep can become 5 mm. A slot that was positioned from an edge can end up positioned from the center. None of these changes is dramatic, and each one is easy to miss on a glance at the shaded model.
The trouble is what happens downstream. Mating parts are dimensioned against the first version. A drawing may have been checked against it. A colleague may have already approved it. If the second session quietly produces a variant, you now have two near-identical parts and a question nobody wants to answer in a design review: which one is the real one?
Re-prompting also costs time even when it works. Every restatement of the brief is a chance to forget a constraint you added in round three, such as the clearance for a tool or the minimum wall next to a hole. The conversation held that knowledge. The new session does not.
IN PRACTICE
With Leo, our team improves design quality, reduces mistakes, and shortens time-to-market.
- Uriel B., Field Warfare and Survivability Specialist
What drifts when you regenerate
It helps to be specific about which features move, because that tells you what to check. In our experience of reviewing generated parts, the features that shift most often fall into a short list:
Fillet and chamfer sizes, especially small ones that the prompt only mentions in passing.
Hole depth and counterbore steps, where the description gives a diameter but not a depth.
Datum choices, meaning which edge or face a slot or hole is measured from.
Overall proportions, when the prompt gave a ratio or a rough size instead of a number.
Feature order, which affects how easily the part can be edited later.
Each of these is a place where the tool fills a gap. A prompt that states every dimension closes most of the gaps, and our guide to writing prompts that produce manufacturing-ready parts covers how to phrase them. But even a very careful prompt only reduces the room for variation. It does not remove it, and long prompts are tedious to repeat exactly.
There is also a limit on what a prompt can carry. A part that reached its final form through five corrections contains decisions that no single sentence captures. The cleanest record of those decisions is the part itself.
Carrying the part forward instead
The alternative is to stop describing the part and bring the part itself into the next piece of work. Leo AI recently shipped this in its release notes for the week of September 29 to October 4, 2026: you can now bring a CAD part you created in a different conversation into a new one. That wording is the whole claim, and it is worth reading narrowly. It is about opening a part made in an earlier conversation inside a later one, so you start from the geometry you already have.
The practical effect is that the second session starts from the exact part, not from a description of it. If the bracket had a 3 mm fillet, the part you bring in has a 3 mm fillet, because nothing was regenerated. Anything you then ask for, such as a mating plate or a cover, is built around known geometry instead of around a fresh guess.
Leo is an AI assistant for mechanical engineers, trained on more than a million pages of standards, books and articles, and it can connect to a company's own knowledge base. It works as an intelligence layer on top of existing PDM and PLM systems, not a replacement for them, and it offers integrations with leading platforms including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter and Arena PLM, among others. Carrying a part between conversations is a small feature inside that larger picture, but it addresses a very specific frustration.
One caution applies to any tool. A part carried forward is only as good as the first generation. If the original had a problem, carrying it forward preserves the problem. Our look at why most AI-generated CAD models are not ready for production explains what to check before a generated part earns that kind of reuse.
How reuse differs from searching your vault
Carrying a part between conversations is a different job from finding an existing part in your company's library, and it helps to keep the two apart. The first is about continuity inside your own current work: the part you made an hour or a week ago, still in the form you left it. The second is about discovery: a part someone else released years ago that you did not know existed.
Both save time, and they can feed each other. A generated part that has proven itself can be released into the vault, where geometry-aware search can find it later. Our article on why engineers spend so much time redesigning parts that already exist covers the vault side in detail.
The distinction matters for how you evaluate tools. A tool that is good at search may say nothing about what happens to a part between two chat sessions. A tool that carries parts forward may say nothing about your released library. Ask about each separately instead of assuming one implies the other.
It also matters for what you can trust. A released part has been through review, so its geometry is settled. A part carried forward from a conversation has not necessarily been reviewed. Treat it as your working copy until someone has checked it, and release it only after that check.
A simple test for any tool that claims to hold details in place
You do not need a vendor demo to find out whether a tool preserves geometry. A short test on your own work is more convincing, and it takes less than an hour.
Generate a small part with at least three details you can measure: a fillet, a counterbore and a slot are good choices.
Record the actual values from the model, not the values you asked for.
Close the conversation, open a new one and bring the part in using whatever method the tool offers.
Measure the same three details again and compare them to your record.
Ask for one modification, such as a second hole, and check that the original details did not move while the change was made.
The last step is the one most tests skip, and it is the one that matters. A tool can open a part faithfully and still disturb it when asked to change it. If the original details survive both the transfer and the edit, the tool is holding geometry in place. If any of them shifts, you have learned that before it reached a drawing.
It is also worth testing the multi-part case. Holding one bracket steady is a modest task, while keeping several parts consistent with each other is harder, and why text-to-CAD struggles with assemblies explains why that boundary is where many tools run out of road. If your work is mostly assemblies, test there first.
Finally, keep a record. A short note listing which parts were generated, in which conversation, and what was checked costs a minute and spares a long argument later about where a dimension came from.
FAQ
Keep approved geometry intact
See how Leo AI supports engineers designing with AI.
Leo AI is built for mechanical engineers. Bring a part from an earlier conversation into a new one and keep building on the geometry you already have.
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
