Essay

Constitutions, not instructions, for models

Sharing the reasons behind our judgment gives models something that more process alone cannot.

IInstructions

Marks set by hand, one at a time. The rest stand as they came.

9 set by hand

IIA constitution

Three reasons. Every mark takes its direction from them, even where nothing was set.

3 reasons · 1,050 marks

I started thinking about this when I worked on the support agent at Perplexity, where I used to say that context was the center of the universe. It was everyone’s job to keep the model informed, in detail, just as they would need to be if they had to reply themselves, because the model was expanding the reach of our own knowledge and we were still responsible for what it said on our behalf.

That observation has stayed with me, though the question has become a little different: what do we need to communicate for a model to exercise judgment in a way that makes sense to us?

Getting what someone knows out of the way they work, and into a form others can use, is old work. When Denis Diderot was working on the Encyclopédie in the eighteenth century, he found himself in the company of all sorts of experts, immersing himself in their day-to-day and the technical specificities of their métiers, visiting workshops and learning to use their tools himself. Trying to operate the machines meant discovering how difficult it was to get the process exactly right, something the encyclopedia’s preface acknowledges when it describes the editors becoming, “so to speak, an apprentice,” and making bad work themselves in order to teach others how good work is made.1

In his entry on “Art,” which here includes craft and technology, Diderot honored the intelligence of those tools:

In what physical or metaphysical system do we find more intelligence, discernment, and consistency than in the machines for drawing gold or making stockings, and in the frames of the braid-makers, the gauze-makers, the drapers, or the silk workers?2

This is why I wanted to bring Diderot into this, because what interests me in that passage is the possibility of human intelligence taking new forms through the tools we make. I don’t think we fail to appreciate intelligence, we do, and we reward some forms of it quite generously. Doctors, lawyers, engineers, people whose work we readily describe as intellectual, tend to receive a kind of recognition that the person operating machinery for mass production, or working the land, does not. But there is intelligence in that work too, in knowing how something should feel, when to adjust it, how to get a process right, and I think this kinetic, physical aspect of intelligence is particularly easy to undervalue.

I like that Diderot tried to use the machines himself, that he encountered the difficulty rather than only admiring the result. It makes me think about how much of what we know takes shape through doing something, through working with a tool until we can do things with it that we couldn’t do before. And this is what I find exciting about AI, the possibility that something of our own intelligence could take a form we hadn’t imagined, that working with a model could help us express or develop something we wouldn’t have arrived at otherwise.

So much of what someone knows is held in the way they work,3 in things they no longer think to explain, and I don’t think a constitution can capture all of that. What it can do is give us a way to share some of the reasons behind our judgment, the things we care about and why, so the model has something to work from when we haven’t specified every choice. This is what I mean when I say that sharing those reasons can give it something that more process alone cannot, even when that process involves elaborate chains of agents, instructions and workflows.

Anthropic already uses this language for Claude, writing a constitution that generally favors “cultivating good values and judgment over strict rules and decision procedures,” and tries to explain the rules it does keep.4 That document is written for one model by the company that trains it, but I think the idea belongs to anyone who works with one, that we can write constitutions, not only instructions, and give the model a clear point of view and conviction to work from.

I keep thinking about the difference between giving a model information about you and giving it a way to exercise judgment. A memory can supply content, while a constitution can explain why that memory matters, what you attend to, what you find moving, which tensions you want to preserve, what you consider an inadequate resolution. These are reasons it can work from when a choice comes up that you haven’t specified individually, rather than another set of steps to follow.

Receiving and developing a human perspective is part of what I think these models are for, and although authorship still matters to me, independence isn’t necessarily the measure of success. When I commit to sharing my world view, my intrinsic motivations, my intuitions and the things I wouldn’t usually think to share, the outcome can feel like an expansion of my mind rather than a mere automation tool. There is something spiritually fulfilling in that, something that enriches my intellectual life.

I can recognise the sensibility behind a result even though I wouldn’t have arrived at its particular expression through my own work, and it is this continuity within the surprise that interests me, the sense that there is something of me in it alongside something I couldn’t have articulated on my own. But I also want that expansion to include something I can learn from, rather than increasingly persuasive agreement or cognitive surrender, because a model can echo my convictions beautifully without developing their implications very far.

I want it to have room to expose tensions between my commitments, follow an intuition somewhere uncomfortable, or offer a reading I initially resist, which means conviction has to leave room for revision. Otherwise, sharing more could make the collaboration narrower, and I don’t think sharing everything is the answer either. Some small, precise disclosures may carry more than a long autobiography, which leaves me wondering which of the things I’ve considered irrelevant to tell a model might turn out to change the quality of its judgment.

I keep coming back to the workshop, to trying something and noticing what I usually leave unexplained, then finding a way to share the reasons behind it rather than only the steps. That is how I try to work, choosing a few precise things to share, leaving room for the model to disagree with me, revising as I go, and judging the result by whether I learned something rather than whether it agreed with me.

And none of this feels confined to AI. A feed, a camera’s automatic mode, the defaults in the software a team shares, any technology that makes choices on our behalf carries someone’s reasons, stated or not, and writing them down is one way they become something we can question and share. From where I stand, AI has been sold as a replacement for human qualities when it could be a way to communicate, explore and challenge our perspectives and convictions, and I think this is part of why it feels unrelatable to people. Our motivations, intuitions and ways of seeing have a place in this relationship, giving us something to contribute and, through working together, something to discover.

Notes

  1. Jean le Rond d’Alembert, Preliminary Discourse to the Encyclopédie (1751): “Il a donc fallu plusieurs fois se procurer les machines, les construire, mettre la main à l’œuvre, se rendre, pour ainsi dire, apprentif, & faire soi-même de mauvais ouvrages pour apprendre aux autres comment on en fait de bons.” It was often necessary to get hold of the machines, build them, put a hand to the work, become, so to speak, an apprentice, and make bad work oneself to teach others how good work is made. The French text on Wikisource. ↩

  2. Denis Diderot, “Art,” Encyclopédie (1751). Original French text. English wording quoted in Stewart Brand’s Maintenance: Of Everything, “Great Manuals in History,” fifth draft. ↩

  3. The philosopher Michael Polanyi put it as “we can know more than we can tell,” in The Tacit Dimension (1966). ↩

  4. Claude’s Constitution, Anthropic, 2026. ↩