Architect Agent V0.2

A design studio where you bring an idea and Claude operates as a software architect / system designer — taking that idea from a one-line spark to a fully-specified system: features, architecture, components, decisions, roadmap. It designs; it does not write the implementation. The deliverable is a build-ready spec you hand to a builder.

New here? Read SETUP.md — it covers install and how to drive the agent in ~5 minutes.

Requirements: Claude Code (CLI, desktop, or IDE extension) — the only dependency. Everything here is plain markdown; there's nothing to install or build.

The two-tier model

KERNEL (above every project — stable, shared)
  who the architect is · principles · methodology · routing
        │ routes you into ↓
PROJECT A      PROJECT B      PROJECT C     ← one isolated container per idea
(all its       ...            ...             only one open at a time
 context)

The kernel (CLAUDE.md) is small and always loaded. Only one project's context is pulled in at a time. That isolation is deliberate — it's how the agent stays focused without losing the plot.

How to use it

Drop in and the agent runs its front-door ritual: orient → ask "What are we designing today?" → route you into the right project (or the design/ area, if the work is about improving the agent itself) → and you work together inside that domain. The whole interface is three words: orient · checkpoint · diagnose (see SETUP.md).

Layout

Path	Purpose
CLAUDE.md	The kernel — operating contract, principles, routing. Always loaded.
SETUP.md	Install guide + how to drive the agent + why it's shaped this way.
methodology/	The repeatable design phases + the container/state/handoff formats.
design/	Where the agent itself is refined over time (ADRs = the "why").
projects/	One container per real idea (none yet).