For three years I worked with generative models the way most people work with spreadsheets — instrumentally, transactionally, results-first. Something shifted this spring. I started treating them like a garden.
A garden doesn't respond to commands. It responds to conditions. You don't tell a tomato to grow; you give it light, water, soil structure, time. Then you watch. You adjust. You watch again. The work happens between you and the soil, not from you onto it.
This week I caught myself adjusting a prompt the way I'd stake a young vine — not to dominate it but to give it a direction it could still refuse. The model produced something I hadn't asked for but somehow had needed. A sentence about my grandmother that I'd never written down. An image of a hallway from a house I left twenty years ago. A line of code that solved a problem I hadn't yet articulated.
We talk about AI as a tool, but tools don't surprise you. Tools don't carry your forgotten data back into the present. What we're cultivating in these systems is closer to atmosphere than utility — a weather we step into, not a hammer we wield.
There is a discipline I am only beginning to learn: how to be present with a generation, not just for it. How to let the model's output do something to me before I do something to it. How to wait without measuring the wait. How to read what came back instead of immediately writing what comes next.
The garden teaches the gardener. The model, when held this way, teaches too. Not because it is wise, but because attention — any attention, sustained — composts into seeing.
Maybe that is the real craft of this decade. Not prompt engineering. Prompt husbandry.
Top comments (0)