multi-agent-brainstorming

Multi-Agent Brainstorming — Structured Design Review Skill

Skill Overview

Convert a single-agent design into a robust, review-validated design by simulating a formal peer-review process, using multiple constraint-driven agents for sequential design reviews.

Suitable Scenarios

1. Design Validation for High-Risk Projects

When the cost of design decision failure is high—such as for core system architecture, security-related functions, or key components that impact user experience—this skill uses multi-angle reviews to identify potential risks, ensuring the design is thoroughly validated before implementation.

2. Review of Complex System Architecture

For complex systems with multiple interacting components, a single viewpoint may overlook edge cases and dependency relationships. This skill uses specialized roles—such as constraint protectors and user advocates—to comprehensively review non-functional constraints, including performance, scalability, and maintainability.

3. Preventing Premature Convergence in Design

In team discussions or personal reasoning, designs can get prematurely locked onto the first viable solution. This skill forces the questioning of assumptions through a skeptic agent, while a user advocate keeps focus on real user needs, preventing overconfident decisions based on incomplete assumptions.

Core Capabilities

Structured Sequential Review Process

The skill uses a three-stage process: single-agent initial design, sequential reviews by multiple roles, and final integration arbitration. Reviewers intervene in a fixed order (skeptic → constraint protector → user advocate), with each role strictly adhering to its responsibility boundaries to avoid confusion and redundant work common in parallel discussions.

Enforced Separation of Role Responsibilities

Five agents each have their own duties: the lead designer is responsible for producing the design but cannot approve their own work; the skeptic only finds problems and does not propose solutions; the constraint protector focuses on non-functional requirements such as performance and safety; the user advocate represents the final user experience; and the integrator resolves conflicts and makes the final decision. This separation prevents role confusion and review blind spots.

Explicit Decision Log Recording

All design decisions must be recorded: the decision itself, alternative options considered, objections raised, proposed solutions, and the rationale. Any design with incomplete reviews, unresolved objections, or incomplete logs is considered invalid. This mandatory logging ensures traceability, enabling future retrospectives and knowledge transfer.

Common Questions

How does multi-agent brainstorming differ from traditional brainstorming?

Traditional brainstorming typically uses a parallel free-discussion mode, encouraging multiple people to simultaneously propose ideas—often leading to groupthink and false consensus. This skill uses a sequential review mode, where only one agent speaks at a time. Creativity is concentrated in the primary designer, while scoring or critique is distributed across multiple specialized roles. The goal is to expose problems rather than generate more ideas; final decisions must be explicitly logged rather than assumed to converge by default.

When should this skill be used?

Use it when you need higher confidence in the design, lower risk, or a more formal review. This includes: key functions that affect core business, designs involving security or privacy, architectural decisions that may become technical debt, proposals that must be formally reported to stakeholders, and any scenarios where the answer to “If this design fails, do we do everything we can to discover problems early?” is uncertain.

How much time does this review process take?

Time investment depends on design complexity and the number of objections. The initial design stage uses a standard brainstorming skill; each review agent typically needs 1–2 rounds of interaction; the integration arbitration stage depends on the number of unresolved objections. For most medium-sized designs, the full process can be completed in 30–60 minutes. This skill includes a forced termination mechanism to prevent infinite review loops.