Superbrainstorming

Talk through the design with your coding agent, agree on it, then let the agent build it.

Superbrainstorming is a plugin for Claude Code. When you ask your agent to build something, it asks you a few questions first, shows you a design, and waits for your OK. As soon as you approve, it starts building in the same session.

flowchart LR
    ask["You ask for something"] --> design["Agent asks questionsand proposes a design"]
    design --> approve{"You approve?"}
    approve -->|Not yet| design
    approve -->|Yes| build["Agent builds it"]

Why?

Coding agents are good at building things. They can't know what you meant, who it's for, or which trade-offs you care about until someone asks, so a capable agent can still build the wrong thing, and build it well. Superbrainstorming has the agent check with you before it writes any code.

It's a fork of Superpowers by Jesse Vincent and Prime Radiant. Superpowers is a full workflow: brainstorm a design, write a detailed plan, then hand each task to a fresh subagent with test-driven development and code review at every step. That structure helped earlier models stay on track. Current models like Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 can keep an agreed design in context and work through it on their own, so this fork keeps the brainstorming and drops the rest.

How it works

First, the agent makes sure it understands what you're after: what you want to achieve, who it's for, and what "done" looks like. It writes that back to you so you can correct it.

Then it picks one of three paths based on the size of the job, and tells you which one so you can overrule it:

Path	For	What happens
Spike	"Is this even possible?" questions	The agent suggests a quick experiment, you give it a nod, and it reports back what it found. Anything it builds is throwaway.
Bounded	Small, well-scoped changes to existing code	The agent asks what matters and shows you a short design in chat.
Architectural	New projects, new subsystems, big structural changes	The agent asks questions one at a time, suggests two or three approaches, and writes a spec to docs/specs/.

Whichever path it takes, nothing gets built until you approve. Once you do, the agent gets started right away. It tracks its own steps, follows your project's conventions for tests and commits, and checks its work before calling it done. If it finds a problem with the design partway through, it stops and asks you instead of quietly changing course. When it's finished, you get a summary of what it built and how it checked it.

If you'd rather skip the design and just build, say so and the agent will.

Install

In Claude Code, run:

/plugin marketplace add harrymunro/superbrainstorming
/plugin install superbrainstorming@superbrainstorming

Then start a new session.

Already using Superpowers? Disable or uninstall it through /plugin first. Both plugins have a brainstorming skill and a session-start hook, and Superpowers' version moves on to writing a plan instead of building.

Try it out: in a fresh session, send Let's make a react todo list. The agent should load the superbrainstorming:brainstorming skill, say it's treating this as architectural work, and ask what the app is for before writing any code.

Using another agent?