How a designer who can't code and an AI built an operating system together.
"They said Claude Code was just a command-line tool."
We pushed the boundary to see its potential — and what's coming.
Out of the box, Claude knows how to code, reason, and write. But it knows nothing about me — my aesthetic preferences, my project structure, or how I like to work. Every session starts from zero.
The first step was CLAUDE.md — a persistent file that shapes instructions, preferences, and desired behaviors. Design tokens, naming conventions, the Murmur aesthetic, workflow rules. Not a one-time prompt. A living document that models how Chris thinks about design — and travels with every conversation.
Instructions alone have limits. So Chris directed Claude to build tools — a Design Language System for defining taste, a Project Manager for tracking sessions, a Git Visualizer for understanding the workspace. Each tool became a new interface for the AI to learn through.
After weeks of layering — instructions, tools, session logs, memory files — the AI stopped feeling generic. It knew the port registry. It knew the design system. It knew that "make it breathe" means more padding and softer easing. Not "just" a smarter model — but also a personal AI.
The first step wasn't building a solution.
It was building the system that can solve problems at scale.
26 interactive labs, 12+ presets, live theme generation.
26 labsEditorial knowledge base with syntax highlighting and session logs.
Knowledge baseSession dashboard, kanban boards, embedded terminal, live WebSocket.
Real-timeReal-time branch and commit visualization across 30+ repositories.
30+ reposVisual dev server orchestrator. Start, stop, restart — all from one panel.
OrchestratorPresentation builder with live presenter mode and WebSocket sync.
Live modeElectron orbital launcher with process management and web preview.
Desktop appVoice & Vision for spoken input. Sketch canvas for drawing and diagramming.
2 MCPsOut of the box, Claude is trained on billions of tokens. It knows how to code, how to reason, how to write. But it knows nothing about this workspace, this designer, or this aesthetic.
CLAUDE.md files, global instructions, design tokens, port registries. Persistent context that shapes every conversation. Chris defined the Murmur design language here — every tool inherits the same palette, spacing, and motion principles.
Each tool Chris directed Claude to build became a new interface for collaboration. The DLS for defining taste. The Project Manager for tracking sessions. Git Visualizer for understanding the workspace. Tools don't just do things — they teach the AI about the environment.
Within each project, micro-adjustments compound. Session logs, memory files, project-specific CLAUDE.md files. "Make the sidebar breathe more" becomes a learned preference. Each interaction refines the collaboration — not the model, but the shared understanding.
Anthropic believes that Tool Use and Computer Use are the path to AGI — the idea that intelligence isn't just about thinking, but about acting in the world through tools. Chris and Claude lived this thesis. Every tool they built together became a new capability, a new sense, a new way for the AI to understand and shape its environment.
Chris directs, Claude implements. "Build me a Git visualizer." A human with vision, an AI with execution. The first tool is rough — but it works.
The new tool teaches Claude about the workspace. Port registries, session logs, file structures. CLAUDE.md grows organically — not as documentation, but as shared memory.
With richer context, the next tool is more sophisticated. Server Manager can reference Git state. Project Manager knows about running sessions. Each tool inherits the wisdom of its predecessors.
A dense, interconnected network pulsing with energy. Not 8 separate tools — one system that understands itself. The whole becomes greater than the sum of its parts.
| App | Backend | Vite | Notes |
|------------------|---------|-------|-----------------|
| project-manager | 4800 | 5180 | strictPort |
| storytelling | 4850 | 5555 | strictPort |
| gitviz | 4870 | 5174 | strictPort |
| server-manager | 4810 | 5176 | strictPort |
| shell | 4830 | 5215 | Electron |
| workflow | — | 5175 | Vite only |
| dls | — | 5179 | Vite only |
| meetings | 4900 | 5181 | strictPort |
| brain-delegate | 4840 | 5216 | cloudflared |Real excerpt from the project's CLAUDE.md. Structured context that grows organically as the ecosystem expands.
You don't prompt an AI into a personal AI.
You give it tools, goals, and let it grow.
The ecosystem didn't arrive fully formed. It grew in phases — each phase building on the capabilities unlocked by the previous one. Not a roadmap executed, but an emergent trajectory.
Each tool we built made the next one easier — and better.
But more importantly, it made Claude smarter and more personalized for Chris.
The Design Language System isn't just about design — it's a shared language between a human and a machine. This is how Chris teaches Claude about aesthetics and taste, through interactive labs rather than static instructions.
Loading live prototype...
Chris can't code. This is both the constraint and the superpower — taste without implementation ability, direction without distraction by technical details. The collaboration works because of the asymmetry, not despite it.
"The sidebar feels too heavy. Make it breathe more. The transition when it opens should feel like exhaling, not snapping."
padding: 24px → 32px
opacity: 1 → 0.95
width transition: 200ms → 350ms
easing: ease → cubic-bezier(0.16, 1, 0.3, 1)
border: 1px solid → 1px solid rgba(..., 0.06)The translation layer is the collaboration itself.
Compounding intelligence, by human in the loop.