[process] | 2026-07-11 | 11 min

# Brainstorming with AI: Why Clarify Beats Prompt

The value of AI brainstorming is not letting AI think for you—it is using AI to think clearly. Mastering diverge–converge rhythm, separating problem from solution space, and keeping a decision log matters more than picking a brainstorming tool.

[brainstorming][workflow-design][agentic-engineering][hitl]

The same Superpowers brainstorming skill produces a design doc for one team and a premature tech-stack choice for another. Same tool, wildly different outcomes—and the gap is usually not the model. It is whether you treat brainstorming as a learnable process skill.

Mainstream AI development workflows—Superpowers, GStack, gstack, OpenSpec, Spec Kit—all embed some form of brainstorming, clarify, or explore. Industry observation is consistent: the same tooling yields very different results across people. This post is not about “which tool wins.” It is about principles that work across tools: why brainstorming matters more in the AI era, how to collaborate effectively with coding agents, and which anti-patterns let AI decide for you.

1. Why brainstorming matters more than prompting in the AI era

Human–AI collaboration hinges on expressing intent clearly. In practice, people often know a rough direction or goal but cannot articulate it—even when they have a mental picture, structuring it in words is hard.

That is normal. Nobody starts with a complete, crisp statement of intent; progressively understanding “what I actually want” is a reasonable path. Traditionally, brainstorming was a soft skill: asking good questions, converging well, deciding under disagreement. Not everyone is strong at it; not every team has time to practice.

With AI and coding agents, the game changes. Skills and workflows embed brainstorming into agent sessions—Superpowers brainstorming, GStack Office Hour, OpenSpec explore, Spec Kit clarify—all make “clarify before build” repeatable. AI will not think everything for you, but it can:

Probe ambiguities you did not notice
Lay out multiple framings so your preferences surface
Record options and trade-offs as you discuss

In other words: the primary value of AI brainstorming is not “let AI think for me.” It is use AI to think clearly. That aligns with the Research→Plan spine in AI Development Workflows Landscape—straighten intent and problem before Execute.

2. Theory: Double Diamond and problem vs solution space

Double Diamond: two diverge–converge cycles

The UK Design Council Double Diamond splits design innovation into four phases mapped to two thinking modes:

Phase	Thinking mode	What you are doing
Discover	Diverge	Understand the problem—not assume it
Define	Converge	Distill insights into a defined challenge
Develop	Diverge	Explore multiple solutions to the defined problem
Deliver	Converge	Test at small scale, filter, ship

Two diamonds means two diverge–converge cycles. Design Council stresses this is not linear—new information sends you back; early prototypes can be part of discovery. That matches AI workflows where “one clarify round was wrong—run another.”

Problem space vs solution space

Dan Olsen in The Lean Product Playbook separates:

Problem space: user needs, pains, jobs to be done—no concrete product or implementation yet
Solution space: mockups, prototypes, features, tech choices—already a specific commitment

Indi Young adds that deep problem-space understanding takes time and should not be collapsed into solution-space delivery cadence (When & Why to Explore the Problem Space).

Mapped to AI brainstorming:

First diamond (Discover / Define) lives mainly in problem space: what problem, for whom, what counts as success?
Second diamond (Develop / Deliver) enters solution space: architecture, phases, acceptance criteria?

Common failure: discussing “React vs Vue” before the problem is defined. That is solution-space talk without a converged problem-space artifact—the agent guesses intent or decides for you.

Digital.gov’s HCD guide states it plainly: divergent thinking explores possibilities with constraints temporarily suspended; convergent thinking is decision-making—filter, focus, choose. Success depends on deliberately switching modes, not only diverging or only converging.

3. Where brainstorming sits in AI workflows

In the five-step spine from AI Development Workflows Landscape, brainstorming sits at the Research / Plan boundary:

Research (understand system and intent) → Plan (how + proof of done) → Execute → Review → Ship

Brainstorming mainly serves the front half of the first two steps: before plans, specs, or code, converge “what and why” into recorded decisions.

Relationship to common tools—tools are lenses; process is discipline:

Tool / capability	Typical entry	Role in the flow
Superpowers `brainstorming`	skill trigger	General clarification, neutral probing, design sign-off
GStack Office Hour	gear / mode	Product value lens—“is this worth doing?”
OpenSpec explore / propose	`/opsx:propose`	Spec change exploration, problem and delta alignment
Spec Kit `clarify`	CLI / templates	Structured requirements, spec field completion

They are not a mandatory serial pipeline but different framing lenses. Superpowers is general-purpose; GStack Office Hour feels like investor-style value interrogation; OpenSpec targets brownfield change memory; Spec Kit targets greenfield spec scaffolding. Pick a lens based on whether you are in the first or second diamond, problem or solution space—not on “tool rankings.”

If you use the OpenSpec + Superpowers + gstack stack, brainstorming usually runs before or alongside OpenSpec propose: Clarify first, then spec, then Superpowers writing-plans and the TDD execution chain.

4. Operating principles: diverge–converge cycles and decision logs

Regardless of tool, effective AI brainstorming follows the same operating discipline.

1. Every diverge round needs exit criteria

Diverging is not “chat until the agent says we’re good.” Before ending a diverge round, produce at least:

Several candidate framings (how the problem could be defined)
Explicit assumptions (we believe X; not yet validated)
Open questions (unclear items for the next round)

Diverge without exit criteria becomes idle talk—tokens spent, no decisions kept.

2. Convergence must be a human-involved decision

ACM CHI 2025 research finds that in idea selection (convergence), users prefer high human agency—AI should respond to explicit feedback and behavior, not decide unilaterally. That aligns with social loafing / smart loafing risks in Springer BISE 2025 on human–AI brainstorming: outsourcing thinking to AI looks productive while human agency fades.

At convergence you should be able to answer:

What was chosen (one-sentence decision)
What was ruled out (at least 1–2 rejected options)
Why (evidence, constraints, trade-offs)
What remains open (for the next diamond or Execute)

3. Decision log as minimal source of truth

Chat disappears; context windows roll; a new session means amnesia. A decision log is the durable artifact of brainstorming—and input to Plan, Spec, and Execute.

Minimal fields:

Field	Meaning
Decision	Final conclusion for this round (one sentence)
Alternatives considered	Options discussed but not adopted
Rationale	Why this choice; key constraints
Open questions	Assumptions or validations still open
Scope	Problem space or solution space

Write it in OpenSpec design.md, a Superpowers design spec, or docs/decisions/—format matters less than retrievable, citable, auditable. Same principle as AI4SE workshop design strategy (“write decisions back; don’t lose rationale during convergence”), applied to human–agent 1:1 sessions.

4. Allow lazy-loading of intent

You need not load full intent at once. Like lazy loading in the project knowledge layer: converge core problem-space decisions first, expand solution-space detail on demand. Round one only answers what problem, for whom, and what success looks like; round two covers architecture and implementation—progressive clarification is a feature if each round has a decision log.

5. Anti-patterns: four ways AI thinks for you

1. Diverge only, never converge

A pleasant chat with a dozen options and no explicit “we choose A because…” Execute then drifts on the agent’s interpretation.

IEICE Trans. 2025 shows simply adding GenAI to group brainstorming does not automatically boost creativity; human-generated ideas can drop significantly. Presence of the tool ≠ effective process—interactive turn-taking and explicit convergence are the interventions.

2. Fake convergence: AI decides

The agent summarizes “therefore I recommend B,” you nod, but you never engaged the trade-offs. That is smart loafing: clarify looks done; agency stayed with AI.

Test: without the agent, can you tell a colleague in three sentences what you decided, why, and what you rejected? If not, it was fake convergence.

3. Mixing problem and solution space

Before defining the user job, you debate microservices vs monolith. The agent follows solution keywords; the problem definition gets silently replaced.

Practice: until the first diamond closes, no deep tech selection—only problem statement, users, constraints, success criteria. Open solution space in the second diamond.

4. No decision log: chat as burn

Conclusions live only in the session. New agent, new colleague, same ambiguities a week later. The workflow unit should upgrade from prompt to process asset with entry, artifact, and checkpoint—the decision log is that artifact.

6. Minimal action list

Before the next feature or change, you need not switch tools—try these three steps:

One problem-space round: with the agent, only discuss what problem, for whom, and what success looks like; converge to a decision log of at most three items—no code, no framework choice.
Explicit mode switch: tell the agent “entering convergence—list options A/B/C; I will choose”—make convergent thinking an executable instruction.
Persist before Plan: write the decision log to spec / design / docs/decisions/, then enter writing-plans or /opsx:propose.

Brainstorming is not decoration on the AI development workflow. When agents can write code and edit repos, the quality of Clarify caps the ceiling of Execute. Tools will evolve; Superpowers and GStack lenses will change—but diverge–converge rhythm, problem/solution separation, and decision-log discipline transfer across tools.