Workflow Harness — Productivity skill for Claude Code

Workflow harness

A team of Claude Code agents that plans, researches, designs, builds and reviews software. Each agent works by written rules, and hooks enforce those rules as the agent works.

I built this over a year of research and development, and I use it on my own work. I'm sharing it with anyone who wants to get more done with AI agents. Take it, change it, and make it yours. If you want to talk about how it works or how to adapt it, reach me through my GitHub profile, [@jaymeany](https://github.com/jaymeany).

The team has five roles: Orchestrator, Research, Designer, Dev and QA. Each role runs as its own Claude Code session, in its own folder, with its own identity, mandate, protocol, hooks and memory. The roles message each other directly when they need an answer.

You own the decisions. The agents surface them, frame the options and recommend.

The tools in this harness are my answers to problems any agent workflow runs into. They're my personal choices. You may have your own tooling for the same problems, and that's fine.

Problem	My answer
Work has to persist across sessions and agents	Trello. Each card holds one unit of work and its context
Agents need to understand the codebase efficiently	Axon. It indexes the code into a graph agents can query
UI and interaction work has to be judged on the real thing	Storybook. It renders the project's real components, so what an agent sees is what ships
Review has to see the product the way a user does	Playwright MCP. It opens the product in a real browser for screenshots and checks

Trello sits behind an adapter, so swapping it is a new adapter, not a rewrite. See "Use another board" below. Storybook, Axon and Playwright MCP are optional.

Why the code holds up

The quality comes from the structure.

• Work is split by role. Research documents, Design builds the surface, Dev writes the code, and QA reviews it. Each role owns the quality of its own domain.
• **Eve