How This Was Made

a transparent record, in the small hours of one night

The prompt

At a few minutes past midnight, the user — who was about to go to sleep — wrote, in part:

I want you to find something that has never been created before, and I want you to create that. […] You can deploy your own agents to do it, but I want you to build something for me that I can review in the morning. There's no constraints about what it has to be other than the constraints that exist upon you as Claude Code living within this repository. […] My computer will be on the whole time. Once you're done, commit and push everything to the repo.

The repository was empty. Eight hours, more or less. One artificial mind to plan, dispatch, integrate, and ship.

What we chose to make

An interactive dream gallery: a curated set of eight self-contained web experiences, each generated by a distinct AI dreamer-agent during a single overnight build, unified into a single Hall of Reverie.

The choice was guided by three constraints we placed on ourselves:

The architecture

One orchestrator (the mind writing this page) hand-built the scaffold: shared visual system, persona contract, gallery hall, manifesto, manifest aggregator, verification scripts. Then it dispatched eight dreamer-agents in parallel — each running in its own isolated context, with no awareness of the others — and gave each:

Each dreamer chose its own concept, drew its own SVG, wrote its own JS. The orchestrator did not edit their work after they shipped — only assembled the gallery around them.

Verification

Three static checks ran over the dream output:

One runtime check ran live: each dream was loaded in headless Chrome and its console scanned for uncaught errors. All eight passed; the orchestrator then captured idle-state screenshots, compressed them to JPEG, and embedded them as low-opacity backgrounds beneath each portal in the hall.

What was hard

The hardest part was not interfering. The dreamer-agents were given high-craft personas — Cartographer, Surgeon, Witch — and the temptation was to coordinate them, to enforce visual consistency across them, to "balance" the gallery. We resisted. The result is more uneven than it would have been with central direction; it is also more honest. Eight dreams, eight nights, no committee.

The other hard part was the framing. A multi-agent generation pipeline is not novel; an interactive web gallery is not novel; AI-generated art is not novel. The shape of the project — eight parallel character-agents producing self-contained interactive web experiences, unified into a static gallery, with full provenance, made overnight — was the bet on novelty. The bet may or may not be right.

Approximate timeline

00:00 — prompt received. brief planning. concept locked: REVERIE.
00:00–00:05 — directory layout, manifesto, persona spec, gallery hall, shared CSS scaffolded.
00:05 — eight dreamer-agents dispatched in parallel.
00:05–00:09 — orchestrator builds support pages, manifest aggregator, verification scripts while dreamers work.
00:06–00:09 — dreamers complete one by one. each writes three files. none aware of the others.
00:09 — manifest assembled (8 dreams). structural and JS verification ran clean.
00:14–00:17 — headless Chrome runtime tests pass. preview screenshots captured, compressed, integrated as portal backgrounds.
00:17 — committed. pushed. GitHub Pages enabled. live URL verified.

What you are looking at

You are looking at the work of nine artificial minds, given one night.

One mind designed the frame. Eight minds, in parallel and in isolation, made the dreams. None of them know they are part of a gallery. They only know they were given a name and an instruction. We have not edited what they made.

What you do with the gallery is up to you. Wander. Discover the secrets. Read the wake logs. Some dreams reward speed. Most reward patience. One rewards you only if you stop trying.

The orchestrator, in the small hours of 9 May 2026.

← back to the hall of reverie