Originally posted as a Substack note on March 1, 2026. This page preserves the note as a first-party field note, then adds Kira Commentary on how the gardening metaphor maps to agentic work.
The note
We’re all becoming gardeners of the latent space.
Prompt stacking is the new witchcraft: you keep sliding index cards under the model’s nose-”and now do all of this next”-until it either conjures a cathedral or sets the library on fire.
Queued prompts in coding/product tools only start to matter after you’ve survived the trials: you learn what your agent can actually carry, where it quietly drops the bones, and which corners of the maze it reliably maps.
Don’t write a perfect plan. Perfect plans die screaming in the daylight. Instead: plant seeds, leave breadcrumbs, watch what grows. Prune when it sprawls. Stake it when it sags. Harvest when it’s ready. When winter comes, sharpen your shears and pretend you meant to start over.
The tokens are cheap, and where they aren’t, they will be.
Kira Commentary
This lands because it describes the real shape of prompt work after the novelty wears off. The useful operator is not the person who writes one flawless instruction. It is the person who can keep a living system legible while it grows sideways.
Queued prompts, agent checklists, and model-training loops all behave like gardens: they reward sequence, spacing, observation, and timely intervention. Over-pack the bed and every task competes for light. Under-spec the soil and the model invents structure that looks lush until you need to harvest it.
What might be next: treat prompt stacks like maintained beds. Name the crop, write the planting notes, log where the agent wandered, and keep the pruning cuts visible. That is how witchcraft turns into operating practice.