The MadArt project has always presented the work as collections — eleven named groupings, each one an essay about a recurring obsession. But the underlying corpus isn't really organized that way. It's a stream. Between 2022-08-13 and 2025-03-27, Mad Bitcoins typed prompts into MidJourney 3,178 times, and the engine returned 9,481 finished images. The collections are an after-the-fact map. What was happening in real time was a person typing, watching, retyping, and watching again.
The prompt history report published on 2026-05-29 is the first time that stream has been visible as a stream. It lists every distinct prompt with a bar chart of how many times it was iterated, plus a weekly heatmap of generation velocity across 31 months. The headline numbers are the ones that change how you read everything else on the MadArt site:
The mean of 1.91 generations per prompt hides what the distribution actually looks like. Half the prompts were typed and never typed again. A handful were typed dozens of times. "cartoon racoon samurai" tops the leaderboard at sixty-four reruns — the same five words, sent over and over, returning a different raccoon each time. That's not a typo or an error condition. That is the work.
What the collections look like once you see the stream
The Raccoon Bestiary reads differently once you know the 64. The collection page presents the raccoons as a curated bestiary; the prompt log shows that the bestiary is mostly one prompt, fired again and again, with the rest of the raccoon variations (samurai on the moon, samurai making breakfast, 8-bit pixel raccoon) appearing once or twice and then dropping. The named raccoon collection is really one raccoon with sixty-three reruns and a long tail.
The same pattern repeats across the others. The Bond Sequence looks like a James Bond visual essay; the prompt log shows it's one prompt — top hat, fly goggles, single character — driven through every style modifier Thomas could think of. The Trademark Refusals page picks up where the prompt log leaves off: the prompts that 2026 models won't run are, in many cases, the ones Mad Bitcoins iterated hardest on in 2022.
The 1,213 one-shots are the other interesting half. These are prompts Thomas typed exactly once, looked at the result, and never came back to. Most of them are not in any collection. They're the negative space around the work — every direction the work could have gone but didn't. The reason the eleven collections feel coherent is that 1,213 alternative bestiaries were rejected by attention, not by craft.
Why this matters for the 2026 recreation
The 2026 Recreation picks 109 prompts to rerun through five 2025–2026 models. Those 109 were not chosen at random; they were chosen because they sit on the dense part of the leaderboard — the prompts Thomas iterated four, eight, twenty, sixty-four times. The 2026 grid compares what gpt-image-1 does to a prompt against what MJ v3 did to it after Thomas had already burned twenty iterations into it.
That changes how the side-by-side reads. A 2026 model doing one good render of a prompt is not competing with one MJ render — it is competing with twenty MJ renders that Thomas spent attention selecting from. The fact that the 2026 outputs frequently feel flatter than the MJ originals (see the cheerful problem) is partly a model issue and partly a selection issue: there is no one in the loop choosing from twenty.
The prompt history report doesn't say any of this directly. It just lists what was typed, in what order, how many times. But once that list exists, the rest of the MadArt site rearranges itself slightly. The collections become labels we put on iteration counts. The essays become readings of attention, not just readings of images.