Setting up multi-agents for design process
Feb 2026
TLDR — One command, five agents running in parallel: a Product Manager, Researcher, Designer, Content Designer, and Builder to automate and accelerate the design process.
The problem
Early-stage design work is slow not because the thinking is hard, but because the setup is. Before you can explore ideas, you need context. Before you can prototype, you need a direction. Each feels like a prerequisite for the next, so you do them sequentially and the whole thing takes days. AI doesn't fix this by making each step faster. It fixes it by running them at the same time.
The setup
I built a slash command called /design-teams in Claude Code. You give it a project name and a brief. It spawns five agents in parallel, each playing a specific role and each saving its output directly to a project folder. Product Manager writes prd.md. Problem statement, goals, non-goals, requirements, open questions. The kind of doc you'd hand to an engineer. Researcher writes research.md. Competitive landscape, user behavior patterns, design precedents, and 2–3 sharp insights that should shape the work. Designer writes designer.md. 3 distinct design directions, each with a name, core idea, how it works, and what it gives up. Plus a recommendation. Content Designer writes content.md. Voice and tone, microcopy principles, copy for key screens and moments, and sample strings written exactly as they'd appear in the UI. Builder writes prototype/index.html. A working HTML prototype of the primary flow. No build step. Open it in a browser. When all five finish, the coordinator writes a summary.md. One paragraph synthesizing everything. The quality is rough. That's fine. The point is to have something to react to, not something to ship.
Why parallel matters
The standard workflow is sequential: research informs the PRD, the PRD informs design, design informs copy, copy informs the prototype. That chain takes days, and each handoff introduces delay. Parallel agents break the chain. Everything happens at once. The outputs contradict each other sometimes. The designer goes one direction, the PM frames the problem differently. That tension is useful. It surfaces assumptions you'd otherwise make silently.
How it's set up
The whole thing is a markdown file at ~/.claude/commands/design-teams.md. Claude Code reads it as a slash command. No API key, no extra billing. It runs on the same Claude Code subscription. Outputs land in Google Drive, organized by project name.
What's next
I've been running a version of this at work. The personal setup is simpler. No design system integration, no shared tooling. But the structure is the same. The next thing I want to add is a way to feed the outputs back into a second round, where the agents respond to each other instead of just working from the brief.