Why this is built the way it is
The AI is not allowed to touch the numbers
Ask an image model for a cabinet and you will, sooner or later, get a beautiful render with the drawer pulls on the inside. Everyone laughs, and then everyone draws the wrong lesson. The problem is not that the model is bad at furniture. The problem is that generation and fabrication are different problems, and we keep grading one on the rubric of the other.
A generated image has one job: to look right. A build plan has a harsher one: to be right, in a way that gets checked by a table saw with no sense of humour. When Popular Woodworking tested AI-generated furniture plans, their verdict was that the AI understands what a piece should look like but not the physics of how it stands up. That matches my experience exactly, and it is the founding constraint of the tool I'm building.
LLMs guess dimensions the way they guess citations
A language model asked for a cut list produces something that looks precisely like a cut list. The columns are labelled. The numbers are plausible. And each one is a token predicted from context, which means every number carries some probability of being confidently, fluently wrong — the same failure mode as an invented citation, except a hallucinated reference wastes an afternoon and a hallucinated shelf length wastes a board of walnut.
You cannot fine-tune your way out of this, because the failure is structural. "874" is not a fact about language. It is a consequence of geometry: a 900 mm carcass, minus two 19 mm sides, plus two 6 mm dados. Change the dado depth and the number changes with it. A model that produces "874" from statistical association will keep producing it after you change the joinery — and that is the version of wrong that finds you at the saw, blade already spinning.
The split: the model proposes, the kernel disposes
So the architecture draws one hard line. The language model works on the side of the problem where being roughly right is useful: reading your sketch, guessing that you mean a wall-hung bookshelf, suggesting that dadoed shelves suit both the piece and the tools you told it you own, offering a deeper variant when you say it looks flimsy. Design is a conversation, and models are good at conversation.
But the model's output is not drawings and it is not numbers. Its output is a parametric model — components, constraints, joinery choices — handed to a solid modelling kernel, the same species of software inside any serious CAD package. The kernel derives every dimension from the actual geometry. The drawings are projections of that geometry. The cut list is a report over it.
A cut list is a compiler target
The way to think about it: the LLM writes source code, and the kernel compiles it. Nobody asks a compiler for its opinion — given the same source, the machine code is the same, every time, and when the source changes the output changes with it, mechanically. That is precisely the property a build plan needs. Move a shelf, deepen a dado, swap 19 mm ply for 18: every affected number updates because it is recomputed, not re-guessed. The kerf is subtracted because subtraction happened, not because the model has seen many sentences about kerfs.
What this refuses to do
This design gives some things up on purpose. The model cannot "just write out the cut list" — even when it would probably be fine — because probably is the whole disease. It cannot invent a joint the kernel cannot represent. If you ask for something the geometry cannot satisfy, the answer is a constraint error, not a confident number. In a product category that runs on demos, refusing to let the impressive component near the load-bearing arithmetic is a feature that looks like a limitation right up until you are standing at the saw.
It doesn't exist yet, and that's the point of this page
Sketch Me This is being built in the open, evenings and weekends, by one hobby woodworker who got tired of designing pieces twice — once in his head and once in software that fought back. The page at sketchmethis.com describes exactly what it will and won't do, and the early list decides what gets built first. If the architecture above strikes you as either exactly right or importantly wrong, I genuinely want to hear which — that is what the list is for.
Join the early list — 25% off for the first 500
There's also a small, fully working taste of the thesis on this site: a free cut list calculator that nests your parts into boards and subtracts the saw kerf for every cut. No signup, runs entirely in your browser.