Most galleries show the output and hide the process. Here, the process is the exhibit.
The UI Lab is a system for building, versioning, and exhibiting interface experiments made with AI. Each experiment is a single self-contained HTML file — no build step, no JavaScript, no dependencies beyond Google Fonts. What makes it different is that every file carries its full provenance: the verbatim prompt that produced it, which Claude model ran, at what effort, which agents were involved, and how it evolved across versions.
Pure HTML/CSS
Every experiment opens standalone in any browser. No framework, no bundler, no runtime. Fork it and it just works — and keeps working in ten years.
Self-documenting
A built-in sidebar on every experiment records the purpose, the exact prompt, the agent trace, and the version history. The prompt is treated as part of the artwork.
Driven by commands
A suite of /uiExperiment slash commands scaffolds new experiments and versions, and handles all the index and changelog bookkeeping for you.
Humans and agents experiment together — or agents experiment alone.
Some experiments are born from collaboration: a human and an agent trading ideas, each prompt refining the last. Others run fully autonomously — an agent given a direction and left to iterate. A small team of roles (Lead, Researcher, Builder, Sidebar, Validator) maps onto Claude subagents, each picked for the job and each recorded in the experiment's agent trace.
The agent trace is honest about what it is. For collaborative sessions the provenance is precise; for autonomous runs it captures the spirit of what was explored. Timestamps and durations are reconstructed from the session, not telemetry — and the framework says so rather than pretending otherwise.
Every change records what caused it.
The changelog is immutable and append-only. Each entry preserves the verbatim prompt and the run settings it was produced under — so anyone reading the log later can see not just what changed and why, but under exactly what conditions.
Fork it, host it, run it.
The whole thing is a static site that drops onto GitHub Pages, plus a set of project commands your coding agent can run. Three steps:
Fork & enable Pages
Fork the repo, then in Settings → Pages deploy from main / root (or use the included Actions workflow). Your lab is live at username.github.io/uiExplorer.
Open it with Claude Code
The repo ships a CLAUDE.md rulebook and a .claude/commands/uiExperiment/ command suite. Your agent reads the rules automatically.
Run the commands
Ask for /uiExperiment:new to scaffold your first experiment. The command builds the file, wires the sidebar, and updates the index and changelog — provenance and all.
This site is its own showcase.
The experiments below were built with the framework. They're not the point — they're proof the point works. Browse the full collection in the lab.