What the studio believes, written down so we can be argued with.
A manifesto is not a marketing document; it is a contract with ourselves. We rewrite this sheet every year and we keep the old versions so that the drift is visible. This is the August 2025 version, drawn by all four partners and one cat.
Why a labyrinth, and not a maze.
A maze has dead ends; it is designed to mislead the walker so that finding the centre feels clever. A labyrinth has one path; it is designed so that the centre is inevitable but the walking is what changes you. The Greek word for the latter, labyrinthos, predates the maze by a thousand years and meant something closer to "a building with many rooms" than to a puzzle.
We chose the labyrinth as a name because the most interesting thing about generative AI in design is not the optionality — the thousand possible outputs — but the corridor. The path that the model is forced to walk by the rules we have given it. Most studios sell their clients the maze and let them get lost in it. We would rather sell the corridor and walk it together.
Practically: every project at this studio has a single centre that we agree on in the survey room. Every drawing, every constraint, every training run, every model we ship is in service of arriving there. The optionality is a side effect, not the product.
What we hold to be self-evident.
1. Generation is not design.
A model that produces a thousand variations of a chair has not designed a chair. Design is the act of selection under constraint. We build the constraints first and let the generation come last; the inversion of the typical pipeline is the entire point of the studio.
2. Small models, on your hardware.
We train small generative models — usually a few hundred million parameters, sometimes less — on the client's own archive. They run on a single workstation, they fit in a backup, they do not phone home. The largest model we have ever shipped was 1.3B parameters and the client found it too big.
3. The archive is the brief.
We will not train a model for you if you do not have an archive. The archive is what makes the work bespoke; without it we would just be fine-tuning the same foundation model that everyone else uses, and the output would carry the same accent.
4. We sign the work.
Every commission is signed by the partner responsible, on the title sheet of the drawings and in the model card. We do not ship anonymous models. If we got it wrong, our name is on it.
5. The kill switch ships first.
Every model we deliver ships with a kill switch — a documented, single-command way to remove the model, its weights, its outputs, and its traces from the client's infrastructure. We build the kill switch in week one, before we train. If the client cannot turn it off, they do not really own it.
6. No surprise outputs.
A generative model that surprises its operator is a model that has not been constrained enough. Surprise is for clients who like theatre; we are not in the theatre business. We tune until the outputs are predictably interesting and never accidentally embarrassing.
7. Hand-drawn supremacy.
Every important decision in a project is first drawn by hand. The model never gets to make a decision that has not been drawn by a partner. Some of the drawings are bad — that is fine; bad drawings still beat good prompts.
8. Citations always.
If an output of one of our models is recognisably indebted to a specific drawing in the training archive, we name the drawing in the model card. Our clients sometimes find this awkward; we find it unavoidable. The archive is the brief; the brief should be cited.
9. We do not scale.
The studio takes six commissions a year. We have turned down work to keep this true. Growth past this point would force us to systematise the bits we deliberately keep slow.
10. The model is not the deliverable.
The deliverable is the design. The model is a tool that helped us draw it. We are happy if the model is interesting; we are unhappy if it is more interesting than the work.
11. Refusal is part of the job.
We refuse commissions that would teach a model to imitate a living designer's style without consent. We refuse work that exists primarily to replace junior designers. We refuse to train on archives that are not legitimately owned. The list of refusals is part of the studio — arguably the most important part.
The negative space of the studio.
A studio is defined as much by what it will not draw as by what it will. Some of what follows might read as priggish; we have arrived at each line after losing money on a project that violated it.
- Style transfer without consent. We will not train a model on the work of a living designer or studio without their written agreement. Dead designers we will discuss on a case-by-case basis.
- Replacement, framed as augmentation. If the explicit business case for a project is "do the work of three juniors with one senior plus a model", we are not the studio for it.
- Volume-first commerce. Marketplaces that need ten thousand product images by Friday should not be calling us. We are slow, and the slowness is the value.
- Speculative facades. Generative facade studies for buildings the client does not own, in cities the client has not visited, are not architecture.
- Crypto-adjacent provenance. We have no interest in on-chain receipts for design work. Provenance is solved by signing the drawing.
- Foundation-model wrappers. If the proposed engagement is fundamentally a thin layer over a third-party API, we will recommend a software studio instead, and probably name one.
Where do the juniors go.
The hardest question for any generative design studio — harder than the technical questions, much harder than the aesthetic ones — is what happens to the junior designer when the model can render a passable elevation in seven seconds. We have spent a lot of evenings on this question and we do not have a tidy answer. We have a working one.
The working answer is that the studio commits to employing as many designers as the work requires, and that the work is defined as everything that happens before and after the seven-second render. Survey, constraint-drafting, taste, criticism, redrawing, taste again. The seven-second render is the cheapest minute of the project. We hire for the other ninety-nine.
We also commit, in writing, that we will not pitch a project to a client on the basis of "this used to take a team of five; now it takes one plus a model". We will pitch on the basis of "this used to be impossible, and now it is merely difficult". Difficulty is still labour; impossibility was the thing nobody got paid for.
An open invitation
If you run a generative AI studio and you have found a tidier answer to the labour question, we would like to hear it. We host a small dinner in Bath twice a year; partners only, no slides, no recording.
This document is alive.
Manifestos that are written once and never amended become embarrassments. We rewrite this sheet annually, on the studio's birthday, and we keep the previous versions in the archive. Version 01 was written before we had taken our first commission; it is wrong in interesting ways. Version 06 is wrong in ways we cannot yet see.
If you spot a contradiction between this sheet and how we actually behave, tell us. Either the sheet is wrong or we are; in either case we owe the studio an amendment.