← Back to Work

Series

Tortoise vs. Hare

Each entry is the same game built twice, by LLMs alone. The hare gets one shot: a single prompt, and whatever comes back ships frozen. The tortoise gets a many-shot, review-automated pipeline. Play them side by side.

The finding underneath is a boundary. Below it — a program you can express in under a thousand lines — one high-effort prompt doesn't just suffice, it wins: a single context holds the whole artifact, and nothing has to be re-stitched. The pipeline's ceremony is pure overhead down here. Above it — an artifact larger than any single context can hold — the one-shot can't finish at all, and decomposition stops being process and becomes the only thing that keeps millions of tokens coherent.

Above that line, drift is the norm: however good the first prompt, the spec moves — the wording was ambiguous, the real need surfaces only once something runs, someone asks "can we make a tweak?" A one-shot freezes its guesses at the moment of generation; a pipeline treats revision as ordinary input. Each entry measures what that difference costs — not in tokens, but in human attention per change.

So this is a map, not a verdict. Some entries sit below the line and admit the sledgehammer met a thumbtack; others exist to show you the wall — the game one prompt cannot finish, where the model's own instinct to architect first runs out of room mid-sentence. One-shot until you hit the wall, then decompose. Both artifacts stay frozen as their process produced them; the prompts are part of the exhibit.