A self-built Claude Cowork / Claude Code plugin that runs the full Design Thinking process (Empathize through Deliver) as a phase-gated multi-agent system, with a coordinating Overseer, specialized sub-agents for each discipline, and its own continuously-updated knowledge base of UX frameworks and patterns.
EDGA — Experience Designer Generalist Agent started from a simple frustration: every real design engagement runs through the same disciplines (research, strategy, systems, prototyping, testing, delivery), but a single chat with an AI assistant has no memory of which phase you're in, no gate to stop you from skipping straight to pixels before the problem is framed, and no persistent record of what was decided and why.
EDGA is my answer to that: a plugin that turns Claude into a small, disciplined studio. One Overseer agent keeps track of where a project is in the process and hands work to the right specialist; the specialists never freelance outside their phase; and every phase ends with a human decision gate, so nothing moves forward without my explicit sign-off.
The Overseer is the only agent that talks to me directly and the only one with the full project context. It reads intake data and session memory, decides which phase the project is in, injects the relevant context into a sub-agent, and enforces the human gate before advancing. Sub-agents never load project context on their own. They only work with what the Overseer hands them, which keeps each one focused and keeps token usage predictable on long engagements.
Phases are gated: Empathize → Define → Ideate → Prototype → Test → Deliver, each with a documented "artifact contract," a specific output the phase must produce before the gate opens. I approve or send back the work at every gate; the Overseer never advances a phase on its own judgment.
EDGA isn't one-size-fits-all — a quick gut-check and a multi-month retainer need different amounts of ceremony. Onboarding starts with /edga-new, a structured intake across 9 blocks (goals, users, constraints, existing artifacts, timeline, and more) that scopes which mode fits.
Projects running past ~3 months switch to a Cycle Model: the Memory Coordinator manages staleness checks and cycle transitions so institutional knowledge about the project doesn't silently rot between sessions. It also handles exporting that project memory to Obsidian for a durable, searchable record outside the chat.
EDGA's sub-agents reason from a generic knowledge base (UX frameworks, research methods, interaction patterns) that grows with every release. New material only gets absorbed through a deliberately slow, checked pipeline, because a design-thinking assistant is only as trustworthy as the sources it's quoting.
.plugin release with a changelog entry.Running /edga-export-kb mirrors the plugin's knowledge base into a real Obsidian vault. A Home.md note links out to 12 category Maps of Content (MOCs), each holding the notes for that discipline. The export is content-hash based: it only touches notes that actually changed, never overwrites a note I've hand-edited inside Obsidian, and flags anything orphaned.
This vault is the plugin's own generic knowledge: frameworks and patterns any project can draw on. It is not project memory: a client engagement's institutional memory is a separate export, kept scoped to that one project so nothing crosses between engagements.
EDGA is still evolving. Every real engagement teaches it something new about where the gates should be.