The brainstorming Skill for Claude Code: Design Before Any Code Gets Written

Agents are eager. Give one a half-formed feature idea and it will happily scaffold the wrong thing at full speed. The brainstorming skill exists to interrupt exactly that reflex: no code until the design is explicit and you have approved it.

Key takeaways

• brainstorming ships inside obra's Superpowers, the most starred repo in the catalog at 220,998 GitHub stars; the skill itself counts 261,134 installs with 48 added in the last day (skills.sh registry, GitHub, 2026-07-06).
• It runs before any creative implementation work - features, components, behavior changes - and starts by reading your current project context.
• The dialogue is one question at a time, refining intent, requirements, and design until the shape is clear.
• A hard gate blocks implementation skills, code writing, and scaffolding until you explicitly approve the presented design.
• It is one part of the larger Superpowers workflow - the Superpowers guide covers the full set.

What is the brainstorming skill?

brainstorming is a design-first discipline encoded as a skill. When you ask for something creative to be built, it steps in first: reads the project, asks focused questions one at a time, and converges on a written design. Only after you approve does implementation unlock. The value is not the questions themselves - it is that edge cases, API shape, and user flows get decided while they are still cheap to change.

How do you install and trigger it?

From the project root:

npx skills add https://github.com/obra/superpowers --skill brainstorming

It activates automatically before creative work once installed - that is the point - or on request ("let's brainstorm this feature"). The Claude Code skills docs explain load order and project-level installs. One practical note: the skills.sh audit trail flags Superpowers HIGH risk because it hooks deep into the agent's session to enforce its gates. That is the mechanism working as designed, but review it before enabling in unattended pipelines.

How does the approval gate work?

When the shape of the work is clear, brainstorming presents a design and stops. Implementation skills stay blocked until you give explicit approval - a hard gate, not a suggestion. In practice this converts "build me X" sessions from an immediate scaffold into a short dialogue that ends with a design you have actually read. Developers report the gate matters most on features that sound simple; those are the ones agents otherwise start building in the first minute.

How does brainstorming compare to grill-me?

They are complements, not rivals. brainstorming takes a vague idea and shapes it into a design; grill-me by mattpocock (460,658 installs, skills.sh) takes an existing plan and attacks it until it holds up. The natural order is brainstorm, then grill. Adoption says developers want both: they are the two most installed skills in the planning category.

Who gets the most from it?

Developers who delegate real implementation to agents. The failure mode it prevents - confident code built on an unexamined design - scales with how much you let the agent do. It also produces a written design as a side effect, which doubles as documentation. Browse adjacent picks in the docs and planning shortlist.

Common pitfalls

• Approving the design without reading it - the gate only works if the approval means something.
• Using it for mechanical changes; renames and refactors with no design space just get slower.
• Skipping project context: the skill reads the codebase first for a reason, and answering against a stale mental model reintroduces the ambiguity it exists to remove.

The cheapest place to fix a design is before it exists. brainstorming keeps you there until the design is worth building.

Common questions