This article first appeared in Harvard Design Magazine on May 4, 2026.
For the past two years I have been teaching a course at the Harvard Graduate School of Design called "Re-imagining the Archive." The premise is simple: take collections that are supposed to stabilize knowledge and treat them instead as something you can work on, work through, and sometimes work against.
We have been inside the archives of the Museum of Modern Art, the Harvard Art Museums, the Institute of Black Imagination, and the American Museum of Natural History. Our goal is not to visualize these archives cleanly, nor to make them legible faster, but rather to see what happens when you stay with them long enough that their seams start to show — when the gaps emerge as new sources of meaning. We do not treat data visualization as a neutral exercise in creating and communicating understanding. My students and I are pursuing, evaluating, and critiquing rhetorical and aesthetic gestures in the pursuit of knowledge creation through these archives.
At the same time, the ground has been shifting. Large language models and related systems have moved from curiosity to constant presence. Every day there is another invitation in my inbox: come talk about prompts, come talk about the future, come talk about what this replaces. The academy seems to be obsessed with these new tools, which are moving so fast we can barely keep up — if we do at all.
Prompts as sketches, models as material
The conventional framing of AI in design assumes a finished output. You ask for a thing; it produces a thing. But the more interesting move is to treat the prompt as a sketch, and the model's response as raw material to be reworked. Norman Klein has argued that data visualizations are always partial, always rhetorical. Treat the LLM the same way and the entire conversation changes.
The prompt is not a question. It is a sketch — a first attempt at a thought, made visible so we can argue with it.
In studio, we ask students to keep every iteration. Not just the "final" generated image but the seventh failed try, the awkward rephrasing, the dead-end branch. The archive of the process becomes its own subject — a record of how a problem was thought, not just what the thinking produced.
Fluency on tap
One of the genuine gains is fluency. A student who could not previously sketch a credible 16th-century printing-shop layout can now produce a dozen variations in an afternoon. The cost of "what if?" has collapsed. This is a real gift, and we should be honest about how much it changes the studio's pace.
But fluency is not the same as fluency-in-service-of-something. The danger is producing a hundred plausible variations and being unable to choose between them — because the choice was never anchored to a thesis in the first place. Fluency without commitment is a beautiful kind of paralysis.
Showing your work (all of it)
The push toward "showing your work" in critique is older than AI. But these new tools change what showing your work means. The trace is now legible in ways it never was — every revision is timestamped, every prompt is a sentence, every dead end can be re-run. The classroom becomes archaeological.
For students, this is bracing. The illusion of the singular genius gesture is harder to maintain when the steps are this visible. The work becomes more collaborative-with-the-tool, and the discussion shifts from "what did you make?" to "how did you think?"
Misuse as method
The most productive sessions in studio happen when students misuse the tool on purpose. Asking for a Renaissance map of a fictional city. Asking for archival metadata in the voice of a 19th-century cataloger. Asking for the prompt itself, recursively, until the system breaks.
This is not a gimmick. It is the same generative move that artists have always made with tools that were built for something else — the photocopier, the synthesizer, the Web. Misuse is one of the few methods that consistently reveals what a tool is actually doing, as opposed to what it claims to be doing.
AI as something you work through
If I had to compress the semester's argument into one line, it would be this: AI is not a thing you ask, but a medium you work through. The asking is a starting move. The working-through is where the design happens.
This means the questions that matter are not "is the output good?" but "what did this round of prompting teach me about the problem I thought I was solving?" The model becomes a kind of clay — responsive, malleable, but with grain of its own that pushes back.
Staying with it
So here is where the semester ends, and where the next one begins. We are still figuring out what fluency-with-thesis looks like in practice. We are still figuring out what kinds of artifacts are worth archiving. We are still figuring out, frankly, what to teach.
But the framing — stay with it long enough that the seams start to show — has held. That is the discipline I want the next cohort to inherit. Not the latest model. Not the cleverest prompt. The patience to keep working when the easy answer has already been given.
