Myles Lambert

[Myles Lambert @mylesjlambert]

https://t.co/pT3e0A1ZoA

Before After

[Myles Lambert @mylesjlambert]

Not worth trying to use up the last 1%

It'll get part way through something and ask me to buy credits to finish it

[Myles Lambert @mylesjlambert]

When will @cursor_ai Pro be included as part of the @X Premium + subscription?

[Myles Lambert @mylesjlambert]

When can we expect an Auto permissions mode for @grok build? @skcd42

[Myles Lambert @mylesjlambert]

Grok CLI needs an "Auto" permissions mode

[Myles Lambert @mylesjlambert]

Luca loves it

[Myles Lambert @mylesjlambert]

Taking delivery

[Myles Lambert @mylesjlambert]

Damn

[Myles Lambert @mylesjlambert]

Finished the Outbuilding 1.0 model

[Myles Lambert @mylesjlambert]

The robotics revolution will make beautiful architecture affordable and ubiquitous

[Myles Lambert @mylesjlambert]

@maticrobots Have 2. One is 🧈, the other is 🥛

[Myles Lambert @mylesjlambert]

Got a second @maticrobots

Our old Roomba got demoted to the guest house when we got the first Matic and has now been demoted to the trash can

[Myles Lambert @mylesjlambert]

This roof system keeps the beautiful timber frame and tongue-and-groove ceiling fully on display from the interior

All the insulation and weatherproofing is above the structure. Meaning: no drywall, no dropped ceiling, just the wood

Here's each layer, from the inside out:

Ridge beam: a 10×14 timber running the length of the building at the peak. It carries half the roof's weight and transfers it down into the gable walls

Rafters: 4×8 timbers sloping from the ridge beam down to the walls, 22 inches apart. They carry the full roof load including snow, wind, the weight of the roofing itself, and deliver it to the stone

Tongue-and-groove decking: Douglas Fir planks spanning across the rafters. From above it's the structural base of the roof. From below it's the finished ceiling you see when you look up

Vapor membrane: a thin sheet on the warm side that blocks interior moisture from migrating into the roof in winter. In summer it relaxes and lets trapped moisture escape

Insulation: five inches of rock wool in two staggered layers. The offset eliminates any continuous seam where heat could leak through

Breathable membrane: waterproof from above to stop rain. Vapor-permeable from below so moisture can escape upward instead of getting trapped

Counter-battens: wooden strips creating a 1.5-inch air gap between the insulation and the roof surface. Air enters at the eaves, warms, rises, and exits at the ridge. This passive airflow dries out any moisture in the assembly

Plywood nail base: the solid surface that the roofing panels attach to

Underlayment: a synthetic sheet across the full roof slope, with a self-adhesive ice and water shield at the eaves. Backup waterproofing in case anything gets past the metal

Zinc standing seam: the outer weather layer. Develops a blue-grey patina over time, can last 100+ years, and floats on concealed clips so it can expand and contract with temperature changes

Every one of these layers is individually modeled

Most buildings this size get a set of hand drawings at best. The roof is a note on a page, not geometry you can inspect

For this building, every batten, every sheet of plywood, even every membrane

Has been modeled and the software has verified position, size, quantity and

Most importantly, that the moisture control, insulation, and ventilation all work together as a system

Before construction starts or a single order gets placed

[Myles Lambert @mylesjlambert]

@OtivDev Nice. Yea I've found that building out the knowledge base portion so that the agent can model the right thing has taken the most time

[Myles Lambert @mylesjlambert]

I've been building software where an AI coding agent writes Ruby code that generates a verified 3D model of a stone and timber building. Every detail, defined in code, implemented by AI, verified computationally before I start building

I over the last few months I've asked Claude to search the internet for anyone doing something similar. It never found anything

The project: I describe what I want to build. Claude researches the technique, creates reference documentation, implements the geometry, and a verification pipeline checks everything is in the right place

The AI doesn't just generate geometry, it proves the geometry is correct, then fixes what doesn't pass

I asked Claude to search for similar projects more than a dozen times across different conversations. Always said it couldn't find anything

So I tried something different. Instead of asking Claude to search, I asked it to deeply understand my project and write me a search prompt I could give to Grok

Claude read through 25,000 lines of modeler code, over a hundred craft knowledge docs, the full build pipeline, the verification system, and distilled it into a prompt describing exactly what to look for and what to exclude. No AI image generators, no generic copilots, no chatbot Q&A. Only projects where AI writes executable code that produces verified geometry

I gave that prompt to Grok

Grok found a whole ecosystem:

@mfranz_on using Claude Code + build123d for parametric CAD with manufacturing-ready STEP output

@ruben_kostard building ForgeCAD, open-source AI-native code-first CAD

@data_shapes building "Claude Code for Revit" (OkPy Agent)

@BIMOpsStudio connecting Claude agents to live Revit BIM models for geometry and compliance checks — FreeCAD AI workbench turning natural language into executable Python for 3D models

@tom_doerr's vcad, an AI agent system for parametric CAD

@OtivDev's grandpacad using LLMs to generate OpenSCAD models

@yongyuanxi doing end-to-end AI code generation in Rhino/Grasshopper

The space is growing fast. CadQuery, build123d, OpenSCAD, FreeCAD on the code-first side. Revit and Grasshopper on the architecture and BIM side

I'd been building in what I thought was an empty field, but lot's of people had the same insight. AI is great with code, so if you define your design in code, AI can build it

The thing that found them wasn't a web search. It was asking one AI to understand my project deeply enough to describe it in terms another AI could search for

Claude couldn't find these projects on the open web

But Claude understood what made mine distinctive

And Grok leveraged its X data access to find recent posts about similar projects

[Myles Lambert @mylesjlambert]

@mfranz_on It might just be that Claude isn't very strong with web searches when recency matters. Tbh I haven't tried googling it "manually"

[Myles Lambert @mylesjlambert]

There are 28 pieces of lead hidden inside this building, but you'll never see them

Every place timber meets stone gets lined with lead sheet before the wood goes in

16 wall plate pockets, 8 ridge beam pockets, 4 L-shaped wraps running the full length under the plates

Why lead? These joints are sealed permanently once the roof goes on

EPDM lasts 50 years. Lead lasts centuries

When you can't get back to it, you don't compromise on the material

[Myles Lambert @mylesjlambert]

The knowledge base isn't prompt engineering. It's a hundred markdown files: species properties, joint specs, load paths, shrinkage rates, roof assembly layers

When I tell the agent to add floor joists it reads the floor assembly docs first and builds to those specs, not whatever it scraped from training data

[Myles Lambert @mylesjlambert]

I tried getting Claude Code to create birdsmouth cuts on rafters. Just the agent and me, no supporting system. It either placed the cuts in completely wrong spots or didn't create a solid object or both

And after days of iteration Claude still wasn't getting it right

Two things had to exist before this could work:

First, a geometry validation pipeline. The agent builds the model, the pipeline checks every dimension and contact point, and the agent gets structured pass/fail feedback until things actually fit. Got this working in January and it hasn't changed much since

Second, a knowledge base. Over a hundred documents on stone masonry, timber framing, roof assembly, etc. The building science that tells the agent how things should actually be built for my specific building

I can't model a new part of the building until the knowledge base covers that subdomain. Without it the agent just makes something up

It'll look plausible, but it won't be right

The validation pipeline took a couple weeks. The knowledge base has taken months

Everyone's focused on the AI framework. The framework is the easy part

Giving the agent the domain knowledge to build something real, that's what takes time

[Myles Lambert @mylesjlambert]

Very cool work! I built something similar in spirit, code-first parametric modeling with verification, to help me design a stone and timber frame outbuilding I'm starting construction on soon. I started with JavaScript and Three.js in the browser but switched to Ruby so I could use SketchUp's native boolean operations and Ruby console. My project is more about getting my building design right than building out a general-purpose tool and UI like you have. Really impressive what you've built here