AI + 3D PRINTING // LESSONS FROM THE BENCH
AI-Assisted 3D Design: Lessons From My Workshop
Small test prints, useful notes and better questions for my AI assistant.
Failed prints, parts that almost fitted, and why I started writing down what actually worked. The original starter pack is still here, along with what I’d do differently now.
ORIGINAL DOWNLOAD // PRESERVED
3D Project Starter Pack V1.0
V1.0 — Explicit Workflow / Legacy & Smaller-Model Edition
The ZIP itself was packaged on 17 August 2026. The workflow inside it had been growing for weeks and months before that through actual projects, failed prints, repeated conversations and a lot of “hang on, didn’t we already learn this?” moments. I have kept the archive unchanged because it is a genuine snapshot of where that accumulated process had reached — not a rewritten “modernised” version pretending we knew the future.
aa871a4d1a2c48756697437172867f3038d05b92524dbd55bf6c4ddbcb373a02“Can you make me a bracket?” became “Can you make an STL?” Then, after a few evenings and a growing pile of colourful failures: “Why doesn’t this bloody thing fit?”
MakerMouse, my AI assistant, could help draw shapes and write the code for a model. The harder part was remembering what we’d changed, what I’d printed and which apparently clever idea had already failed.
HOW THE PACK STARTED
A mate wanted a head start.
The notes grew through real projects: make something, print it, find a problem, ask for a change, then discover another part had changed too. I kept adding reminders about the printer, filament, fits and things we should leave alone.
A mate wanted to try making things with AI without starting from a blank chat. I bundled those notes on 17 August 2026 as the 3D Project Starter Pack V1.0. The archive came later; the mistakes that taught us came first.
I hoped it could save him some of the setup and repeated conversations. It couldn’t promise my results on his printer. It was a starting point, not a magic ZIP.
THE PROBLEM I KEPT REPEATING
Fix one bit. Keep the bit that already worked.
I would ask for a small change and get a freshly improved version of half the project. A connector that fitted yesterday might be different today. That is a very efficient way to buy the same lesson twice.
My useful rule became: if the frame and connector work but the wall mount doesn’t, change the wall mount. Keep the fit that already worked unless there is a clear reason it also needs changing.
WHAT’S IN V1?
Sixteen small files to help the next conversation.
The original files use some fairly grand names. What I wanted from them was much simpler:
Make the thing useful, say what is guessed and keep the parts that already fitted.
Record the machine, material and settings I actually used.
Save a working clip, rail or magnet hole so we don’t start over next time.
Try the tiny questionable bit before spending hours printing the whole object.
Know which file made the part on the bench.
Keep the failed attempt, the successful one and the reason for the change.
The archive also has a README, an optional persona, starting prompts, feedback forms and a release checklist. Its original file names and contents stay as they were. My fuller development notes stay in the build record.
THE RULE I STILL NEED
The printer gets the final vote.
A 0.20 mm gap can look sensible on screen and still come off my printer stuck together. Filament, the first layer and which way the part prints all matter. A gap that worked in a tiny test piece may behave differently along a whole sliding lid.
The UNO case taught me this with a broken latch and three rail gaps. SPENDIOSA taught me that keeping a card secure is only half the job if I can’t get it back out.
SO... IS V1 OBSOLETE?
No. Parts of it are.
That distinction is the reason I want the original pack left intact rather than quietly rewriting it.
The tools have changed since I packaged V1. OpenAI’s published BenchCAD comparison reports 95.9% geometric overlap for GPT-6 Astra versus 83.3% for GPT-5.6 Sol. BenchCAD measures reconstruction of 3D objects from multi-view renders by generating CAD code. That is a substantial improvement in the thing the benchmark actually measures.
OpenAI has also published updated guidance for Astra-era skills: shorter descriptions, less bloated context, progressive disclosure, and fewer elaborate step-by-step recipes where the model can already infer the work.
In my own projects, I’ve found it more useful to give the assistant the notes it needs for the change in front of us. I don’t want a small adjustment to turn into another whole-workshop briefing.
Keep the useful rules; spend less time reciting the whole manual.
WHAT I’D KEEP
A shorter set of reminders.
- Keep the fit that worked. Save its measurements and the print that proved it.
- Change one thing at a time. Otherwise a failed print tells me very little.
- Print a small test piece. A test coupon is just that: the little bit needed to answer the question.
- Label the files and parts. Future Me cannot reliably identify “the good red one”.
- Say what still needs trying. Working once doesn’t tell me how it will wear or how it will print on someone else’s machine.
WHAT I’D CHANGE NEXT
Less paperwork. Better memory.
I don’t want V2 to be an even bigger pile of instructions. I want a few standing rules, the right help for the job and notes that keep growing as I learn.
The standing reminders
Keep what worked. Explain guesses. Don’t quietly replace a file I’ve already shared.
Help when I need it
Bring in the relevant tool or saved skill when the job needs it, rather than reciting every manual first.
The workshop memory
BuildBook keeps the sizes, settings, pictures, failed attempts and results together.
That is the direction I want to take. It isn’t a claim that a finished V2 starter pack is available now.
I STILL HAVE THINGS TO LEARN
CAD makes more sense when I have a real problem.
I’m learning the drawing and modelling tools as I go. Knowing which faces touch, how two pieces go together and where a screw needs to sit helps me ask a better question and spot an answer that looks wrong.
Making a wall thicker isn’t automatically a strength fix. The shape, how it bends and the direction of the printed layers matter too. Those choices need physical tests; a render cannot tell me how the part will bend.
Code-based tools such as OpenSCAD and CadQuery can rebuild a model from sizes written in the code. Other CAD tools let me work directly with the shape. I’m interested in using whichever helps the job; I don’t need one magic winner.
START SMALL
Make the bracket before building an AI factory.
I wouldn’t tell someone starting out to install a pile of servers and tools before making their first part. Start with the idea, learn enough to inspect the model, print something small and tell the assistant what happened.
Add a size-changing model, reusable parts or automation when a project gives you a reason. Whether the assistant runs online or on your own computer, the useful notes are still about what actually worked on your bench.
The tools will keep changing. Keeping track of what actually worked will still matter.
HISTORICAL IMPLEMENTATION
Download 3D Project Starter Pack V1.0
V1.0 — Explicit Workflow / Legacy & Smaller-Model Edition
The archive is intentionally unchanged. It was packaged on 17 August 2026, but the habits inside it had been developing through real projects for quite a while before that. Some setup guidance names the model that was current when it was bundled. Keep that context: this is both a usable starter pack and a snapshot of the workshop at that point in its evolution.
V2 will be published separately rather than overwriting this baseline.
SOURCES + CONTEXT
The two external references behind the Astra comparison.
This article is mostly a record of my own workflow and workshop lessons. The Astra benchmark and prompting guidance come from OpenAI’s September 2026 material:
- OpenAI — GPT-6 Astra: A new generation of intelligence ↗ — BenchCAD definition and published 95.9% vs 83.3% comparison.
- OpenAI Developers — Rethinking skills and prompts for GPT-6 Astra ↗ — shorter skill descriptions, progressive disclosure and less recipe-like instruction.
Benchmarks tell us something useful about model capability. I still need to print the part and try it.
I wanted to make useful things and keep the lessons from each print. Small tests and better notes helped me make the next decision.
Print a smaller test piece. Change one thing. Record what happened. Try again.
At the very least, the plastic spaghetti should teach us something.