Bespoke, not prompted.
This is a long-form note on a position that the studio has held since founding and has, over the past year, refined into a working rule. We will not take a brief that begins with a prompt. The note explains what we mean by that, why we have come to mean it, and what we ask the client to bring instead.
Two words, one trap.
"Bespoke" and "prompted" have been quietly fighting each other in the design-AI literature for two years now, and most studios have not noticed. The fight matters because the words point at incompatible practices that look, from outside the studio door, very similar.
A prompted piece of work is one where the input is a sentence in natural language and the output is one of a large optionality space the foundation model has already memorised. The skill is in the prompting; the artefact is the model's reading of a sentence. The dialect is the foundation model's dialect.
A bespoke piece of work is one where the input is a corpus — a client's archive, their constraint language, their house dialect — and the output is in a dialect that did not exist before training. The skill is in the corpus selection, the constraint geometry, and the training pipeline; the artefact is your house, said back to you. The dialect is yours.
The accent problem.
The reason this studio has come to refuse the prompt-first brief is observational, not theoretical. Across the eleven commissions we ran in 2024 (we were doing this work before the studio was incorporated), the four that produced artefacts the clients used in production were the four where the brief did not arrive as a prompt. The seven that did not produce production artefacts arrived with a sentence the client had already written, hoping that the model would do the rest of the work for them.
The accent problem is not subtle. A foundation image-generator has read so many designers' work that it has an accent of its own — a flat, slightly over-considered visual register that any working designer can pick out at thirty paces. That accent has a name in the studio: we call it "the diffusion shadow". The shadow falls across any output that began with a prompt and was not anchored, structurally, to a client-specific corpus.
You cannot prompt your way out of the diffusion shadow. We have tried. The studio in 2023 tried for six months and we were honest with ourselves about the failure. The shadow falls on the prompted output because the prompted output is, in latent-space terms, an interpolation across the foundation model's training distribution — not a sample from a client's distinct distribution.
Three things on the survey-room table.
When a client writes to us, the first reply — written by the partner on rota that week — asks for three things. Not a prompt. Three artefacts that, between them, define the corpus we will work with.
1. An archive. Whatever exists of the client's prior output that is fit to train into. For a publisher, the proofs and the digital archive. For a foundry, the catalogue of historical castings. For an architectural atelier, the drawings cabinet. The archive does not need to be perfect — it needs to exist and to be releasable to the studio under licence.
2. A constraint sheet. A written list of what the client cannot do, will not do, or believes the work should not become. The constraint sheet is harder to produce than the archive because it requires the client to articulate, often for the first time, the geometry of their house language. We help write it in the survey week; we do not write it for them.
3. A centre. One sentence describing where the corridor leads. Not a brief; not a deliverable; the single point of the work, written in plain language, that the client and the studio can both agree on as the destination. The centre on a Loom file might be "a system that lets our design lead make marks faster without losing the house signature". The centre on the Constraint Canvas file was "help our designers explore, without ever showing them something we cannot cast".
What happened on the Identity Loom survey.
I (Wren) went into the Identity Loom survey week prepared, on the studio's instructions, to refuse to take the file if the client could not produce a constraint sheet. The publisher's editorial director had warned me in the introductory email that they were not sure they could articulate the constraints — they knew when something looked "wrong" for their house, but they could not, on demand, say why.
The first morning of the survey, we did not write a constraint sheet. We sat in the compositor's room for four hours and looked at proofs together. The director would say, of a 1962 masthead, "that is us"; of a 1991 masthead, "that is us with a flu"; of a 2017 masthead from a guest art director, "that is not us at all, that is somebody else's house". By lunch the constraint sheet was writing itself in my notebook in a vocabulary the publisher and I had jointly invented in the morning.
The lesson, which I have written into the studio's intake procedure, is that the constraint sheet is rarely something the client already has. It is something the survey week produces. The thing the studio insists on, at intake, is the willingness to sit in the room and write it; not the document itself.
What if the client does not have an archive?
This is the question we receive most often after publishing positions like this one, and the answer is honest: the studio does not take those commissions. Not because we think such commissions are unworthy, but because the work would not be bespoke. We would be fine-tuning a foundation model and selling the client back the diffusion shadow with a slightly different filter; we would carry the studio's reputation against a piece of work that was, in fact, the foundation model's work.
A new house, with no archive, has every reason to use the prompted tools available to them. They are excellent tools. They are not what we make. The studio's position is that a corridor without an archive is a corridor without a centre, and we would rather decline the brief than walk one with the client.
One paragraph for the studio door.
We do not take a brief that begins with a prompt. We will take a brief that begins with an archive, a constraint sheet, and a centre — in any order, and in any state of finish. The survey week exists to take whatever the client brings and to produce the corpus we can train into. The artefact we deliver is your house, said back to you, in a dialect the foundation models do not have. That is what bespoke means in this studio, and it is the only definition we are willing to write our partners' names against.