These are working notes on what I have been calling jig thinking: a way of understanding what happens when artists and designers build their own tools, and in particular what happens when AI makes tool-building fast, specific and disposable.

The core claim is simple. The tools an artist or designer uses are not neutral. They shape the work. For most of the last fifty years, making meant using software someone else built, universal tools made for millions, with all the assumptions that implies. AI changes that. A working tool can now be made in an afternoon, for one project, one question, one hand. This is a structural shift in what software is, and therefore in what creative practice can be. A jig is my name for the kind of tool this shift produces.

The notes start with the jig itself. They then work outward to what surrounds it: a short history of deskilling, what I see in the classroom, an old argument about form and matter, the five capacities I think a studio education trains, and play.

What is a jig

In a woodshop, a jig is a wooden guide a carpenter assembles to make one cut reliable. It is not sold. It is not reusable across projects in any deep sense. It is scrap that has been shaped to hold a piece in place, or to index a router against an edge, or to space a series of dowels correctly. It is made from the same materials as the thing being built, by the person building it, in the middle of the work.

In the studio, a jig is a small piece of software that makes one creative move repeatable, or one material tractable, or one question answerable. It is written by the artist. It is often finished in an afternoon. It is used, and then retired.

Most of what gets called a tool in creative practice is actually a product. Photoshop is a product. Maya is a product. Figma is a product. These are artifacts made by companies and rented to millions of users, who must learn the product's assumptions to get anything done. A jig is a different kind of object. It is built rather than rented, specific rather than universal, and it reflects the person using it, because that person made it.

The models underneath today's AI tools are black boxes. Their behavior is spread across billions of weights and can only be studied from the outside. A jig is the transparent layer a maker builds around that opaque core. You may not know how the model works, but you know exactly what your jig does, because you wrote it. It gives you a hold on the mechanism of a tool, where a product only offers its interface.

The jig is the object. Jig thinking is the orientation, and the orientation travels even when no jig gets built.

A more traditional analogy: gesso

Every painter gessos a canvas. That is the industry-standard preparation: a universal ground you buy from a manufacturer and apply in a prescribed way. A jig is not gesso. A jig is the moment a painter decides to mix their own ground, a watercolour-paper putty, a paste made with rabbit-skin glue and a specific pigment, a surface that absorbs differently than anything you can buy. The painter does this not to sell a new gesso to the world, but because this particular question about how paint behaves on a surface cannot be answered with the standard preparation.

The key move is understanding the material. A painter who mixes their own ground knows what paint does on that ground in a way no one else does. A programmer who builds their own paint application knows what pixels do in that application in a way no one else does. The jig is not just a tool. It is a form of material knowledge, held in the tool, available to the person who made it.

What a jig is not

A jig is not a product. It is not built to be sold, licensed or scaled. The moment a jig is polished for an audience beyond its maker, it has become something else. This is also what separates jig thinking from most Silicon Valley maker discourse, where the tool is always on its way to becoming a startup.

A jig is not universal. It does one thing, for one question. The constraint is the point.

A jig is not a prompt. Typing into someone else's model is input to someone else's tool. A jig is a thing the artist makes. One keeps the artist in the role of customer; the other makes them the author of the tool as well as the work.

A jig is not permanent infrastructure. It is built for a project and retired when the project ends. That keeps jigs light and specific.

A jig is not a replacement for craft. It is built out of craft and material knowledge. The painter who builds their own ground still has to know how paint behaves. The jig extends the artist's position; it does not substitute for it.

A jig is not quite a harness. Some people now use the term harness coding for the scaffolding built around a model: eval loops, test rigs, the structure that lets you iterate on an agent. The two are close relatives. A jig holds the material at the right angle. A harness holds the agent, something with a will of its own. Both exist to resist the same alternative, which is to receive the output, accept it and move on. The choice in front of every maker now is to steer or be steered.

Deskilling, again

Every new image technology has taken a hard-won skill and made it cheap. Photography deskilled portraiture. Offset printing deskilled printmaking. Photoshop deskilled retouching. Each wave was a real loss, and each opened new ground. The new ground was never automatic. People built new practices and new training to occupy it.

AI follows the same pattern. Some skills get cheaper: technical flawlessness, surface polish, virtuosity as a signal, fluency without a point of view. Others become more valuable than they have ever been: knowing what is worth making, recognizing when something is wrong, having a point of view, framing the right question, staying with difficulty.

The danger I worry about is students skipping the training because the output looks good enough without it. That AI can do the technical work matters much less.

What I see in the classroom

I have been open with students about AI in my classes, and they have been open with me about how they use it. A few patterns keep showing up.

The 90% problem. A prompt gets you to ninety percent, fast, and often the result looks legitimately good. The output is monolithic, though. Asked to iterate, to take it laterally, to change one thing, students can't. They start over and prompt again. The medium never refuses; it just regenerates. The iteration loop never closes.

The beginner gap. AI makes the beginning easy. The people who go deep with it already have foundations: architecture, structure, judgment. They use AI to multiply what they already have. Beginners get fast first results but can't tell what's wrong, what to keep, what to throw away. They are steered rather than steering. The question for teachers is how to build foundations when the easy on-ramp produces output that looks finished.

Friction as a diagnostic. When I learned to code, syntax was never the point. The point was the pain of translating an intention into formal logic, and having the machine show you exactly where your thinking broke down. AI removes that friction, and it feels like help. The friction was the diagnostic tool.

Where ideas come from

Samuel Franklin's The Cult of Creativity (University of Chicago Press, 2023) makes a case I find useful: creativity, as we use the word, is a twentieth-century American invention. It grew out of Cold War anxieties, corporate R&D, the redefinition of American labor, and the same psychometric tradition that produced IQ and personality testing. Before the word, there was training, craft, judgment and command of materials. Most of what an art school teaches is older than creativity.

That history leaves us with two models of where ideas come from.

In the first, ideas come from inside the human. This is the romantic, expressive model: the inner gift, the brainstorm, self-actualization. The creative person is someone who has more of it.

In the second, ideas come from working with materials. The hand learns. The material refuses. The frame gets built while the work moves.

Prompt-in, image-out is a machine built on the first model. The prompt is a wish, and the image is the wish granted. On the second model, AI is another material to be in dialogue with, one with its own grain, and the capacities that have always mattered in the studio reassert themselves. Studio education has always argued for the second model. Jig thinking is that argument applied to software.

Form and matter

Underneath the first model sits an older idea, which philosophers call hylomorphism: form imposed on matter. The maker has the form in mind, and the matter receives it. There is no resistance, no discovery, no surprise, and no new knowledge, because the knowledge is assumed to exist in the maker's head before the making starts. It has been the dominant way of thinking about making since Aristotle.

AI as it is currently deployed is the hylomorphic model on steroids. Describe the form, and the matter arrives already shaped.

There are other traditions. Tim Ingold describes making as correspondence with materials: the carpenter splits wood along the grain, following what the material does, rather than imposing a pre-formed idea on inert stuff. Frei Otto and Antoni Gaudí practiced form-finding, letting hanging chains and soap films compute the shapes of their structures. Philosophies of becoming, from Whitehead to Deleuze, treat form as something that emerges in a process.

The jig belongs to these traditions. It is made from the same scrap as the work, in the middle of the work, in response to what the material is doing. It is how correspondence looks in practice.

Five capacities

If AI takes over much of the technical execution, the question becomes what a studio education actually trains. My current answer is five capacities: sensibility, intention, iteration, critique and play. AI doesn't make them obsolete. It makes them more valuable, but only if we are deliberate about training them.

Sensibility is where work begins. It is an attunement, knowing what's good, being drawn to something before you can say why. Research and formal search live here.

Intention is sensibility condensing into an aim. I have come to think of intention as a heading rather than a destination: knowing what direction to set out in without knowing the outcome. Sailors call the method dead reckoning. You start from where you are, move, take a fix and correct. Intuition reads the fixes and tells you when to hold course and when to change it.

Iteration is the motivated return. Revision is an operation, and AI performs it well. Iteration in the studio sense is driven by the maker's own dissatisfaction, and the standard being judged against develops across the cycles. AI iterates only on command; the wanting isn't there. This is the 90% problem seen from the other side. The loop never closes because the wanting was never the student's.

Critique wraps the cycle. It means stepping outside the work to judge it, usually in company, talking about your work and other people's. What critique produces flows back into sensibility.

The first four form a line of process, and the line is really a circle. Each pass around it develops all four.

Play sits off the line. It breaks the rules of the current circle, asks a different question, and changes sensibility itself. If the four capacities improve the work, play changes what work you are doing. It is also the one that can't be turned into a procedure.

Tools change every six months. These capacities compound over a lifetime.

Play

Play is a relationship to not-knowing. It is the willingness to stay in the unresolved long enough for something to emerge. Planning and problem-solving have their place, and play is a different activity from both.

AI systems, as they work today, generate, optimize and sample. They have no prior expectation that could be violated, so they cannot be surprised by themselves. What looks like creativity in a language model is pattern-completion inside a distribution, which is structurally the opposite of how play works. I hold this claim in the present tense. It describes these systems as they are built now.

Long before AI, the hardest problem in my studios was making sure the work ended up somewhere different from where it set out. I never want students to know where they are going to get to. Setting a goal, mapping the route and arriving is a strong way of working, and it is exactly what a predictive system does well. So we train the other thing: start from where you are, and end somewhere you couldn't have named at the start. A fixed goal ends the not-knowing before you begin. An open one keeps you in it the whole way.

The probable and the improbable

In Towards a Philosophy of Photography (1983), Vilém Flusser argues that every apparatus runs a program, a finite space of probable outputs. A camera used unthinkingly makes the probable image: the selfie, the vacation shot, the family portrait, the default the apparatus wants. Flusser calls the person who executes the program a functionary. The artist's job has always been to find the improbable image, the one the apparatus didn't expect to make, inside its own space of possibilities.

This isn't new with AI. The apparatus changed. The job is the same.

Steering is the name for that job. The functionary is steered by the program and produces the probable. The artist steers the apparatus against its program. In information-theoretic terms the probable carries no information, so steering and producing something new are the same act described twice.

AI adds a twist. The improbable doesn't stay improbable. When an artist finds something new, the program absorbs it: the interesting image gets copied, becomes a convention, becomes a preset, becomes the default the tool offers everyone else. Flusser saw this as the program's slow expansion. With AI the loop is automated. Whatever gets made and posted is scraped into the next training run, and today's improbable image raises the probability of its own kind in tomorrow's model. The artist's position is a moving frontier, and it has to be found again continuously.

None of this is an argument against the probable. The probable is legitimate material inside a workflow: scaffolding, the boring parts, getting faster to the interesting question. The line sits at the level of the whole practice. A workflow that only ever produces the probable is the end of art and design. Use the probable. Don't become it.

A jig is an apparatus you build yourself to push past the probable.

Gap-makers: Harold Cohen and AARON

Harold Cohen began building AARON, a drawing program, around 1971 and kept developing it until his death in 2016. Early on he described AARON as a program he had written. Late in life he called it his other self, and he valued it as a gap-maker, something that provoked his thinking rather than replacing it. He hand-colored AARON's drawings for years, because he was better at color than the program was.

AARON made the gaps. Cohen jumped them. It is a model of collaboration through difference, and the more capable machines get at execution, the more valuable that surprise-yourself mode becomes. Jigs are how we jump the gaps.

What jig thinking holds

Put plainly, here is what I want from AI in my own work. I don't need it to generate the image for me. I need it to help me build a workflow, or a custom paintbrush for this particular image, and then I throw that tool away. Software for an audience of one.

That small move gathers up the rest of these notes. It answers the 90% problem, because you work the apparatus instead of accepting the output. It keeps the maker inside the second model, working with materials, where the tool you shape shapes your work. It exercises all five capacities in the building. It is correspondence in practice. And it is the apparatus you build yourself to reach the improbable.

Designers who build

The wider discourse has started to arrive at the same place from the other direction. "Vibe coding" entered the language in 2025 and was Collins Dictionary's word of the year. Software for one is now a newspaper headline. Industry has independently discovered the audience of one, without the artistic and material lineage that would help it hold the idea well.

For twenty years, software was made by a triangle of engineers, designers and product managers. AI is collapsing that triangle, and the roles are folding into each other. The person best positioned in that collapse is someone who can hold the question, the form and the implementation at once. I call that person a designer who builds: a designer's sensibility with an engineer's fluency. That is what jig thinking trains.

Open questions

This is a living document, and some things are unresolved.

Intuition keeps looking like the more fundamental half of the intention pair. I haven't settled whether it deserves its own place on the list.

The larger questions are about education. When technical execution is no longer the bottleneck, what are we deliberately training for? The studio has one answer, built on contact hours, small cohorts, visible process and crit. Which of those can travel beyond the art school, and what would a university look like if it took them seriously?