Brainstorm

Unified brainstorming skill combining interactive framework generation, multi-role parallel analysis, and cross-role synthesis into a single entry point with two operational modes.

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                    /brainstorm                                │
│         Unified Entry Point + Interactive Routing             │
└───────────────────────┬─────────────────────────────────────┘
                        │
              ┌─────────┴─────────┐
              ↓                   ↓
    ┌─────────────────┐  ┌──────────────────┐
    │   Auto Mode     │  │ Single Role Mode │
    │  (自动模式)      │  │ (单角色分析模式)   │
    └────────┬────────┘  └────────┬─────────┘
             │                    │
    ┌────────┼────────┐          │
    ↓        ↓        ↓          ↓
 Phase 2  Phase 3  Phase 4    Phase 3
Artifacts  N×Role  Synthesis  1×Role
 (7步)    Analysis  (8步)    Analysis
           并行               (4步)

Data Flow:

Auto Mode:
  Phase 2 (artifacts) → guidance-specification.md + selected_roles[]
    → Phase 3 (N × role-analysis) → {role}/analysis*.md (immutable)
      → Phase 4 (synthesis) → feature-specs/ + feature-index.json + synthesis-changelog.md

Single Role Mode:
  Phase 3 (1 × role-analysis) → {role}/analysis*.md

Key Design Principles

1. Dual-Mode Routing: Interactive mode selection via AskUserQuestion, with parameter-based auto-detection
2. Progressive Phase Loading: Phase files loaded on-demand via

   Ref:

   markers, not all at once
3. Task Attachment/Collapse: Sub-tasks attached during phase execution, collapsed after completion
4. Session Continuity: All phases share session state via workflow-session.json
5. Auto-Continue Execution: Phases chain automatically without user intervention between them
6. SPEC.md Quality Alignment: Guidance specification and role analysis follow SPEC.md standards (Concepts & Terminology, Non-Goals, Data Model, State Machine, RFC 2119 constraints)
7. Template-Driven Analysis: Role-specific templates (e.g., system-architect) ensure consistent, high-quality outputs
8. Quality Gates: Automated validation of guidance specification and role analysis against quality standards

Auto Mode

When

--yes

-y

Execution Flow

Phase 1: Mode Detection & Interactive Routing

Ref: phases/01-mode-routing.md

Parse arguments, detect mode from flags/parameters, or ask user via AskUserQuestion.

Mode Detection Rules:

1. If

   --yes

   or

   -y

   flag present → Auto Mode (no question asked)
2. If first arg matches a known role name → Single Role Mode (no question asked)
3. If

   --session

   flag present without role name → Ask user
4. Otherwise → Ask user via AskUserQuestion

Output:

execution_mode

Auto Mode Execution (execution_mode = "auto")

Phase 1.5: Terminology & Boundary Definition

• Extract 5-10 core domain terms from user input and Phase 1 context
• Generate terminology table (term, definition, aliases, category)
• Collect Non-Goals via AskUserQuestion (明确排除的范围)
• Store to

  session.terminology

  and

  session.non_goals

Phase 2: Interactive Framework Generation

Ref: phases/02-artifacts.md

Seven-phase interactive workflow: Context collection → Topic analysis → Role selection → Role questions → Conflict resolution → Final check → Generate specification.

Input: topic description, --count N, --yes flag Output: guidance-specification.md (with Concepts & Terminology, Non-Goals, RFC 2119 constraints), workflow-session.json (selected_roles[], session_id)

TodoWrite: Attach 7 sub-tasks (Phase 0-5), execute sequentially, collapse on completion.

Phase 3: Parallel Role Analysis

Ref: phases/03-role-analysis.md

Execute role analysis for EACH selected role in parallel (excluding ui-designer, which runs sequentially in Phase 3.5).

Input: selected_roles[] from Phase 2, session_id, guidance-specification.md Output: {role}/analysis*.md for each role

Parallel Execution: Launch N role-analysis calls simultaneously (one message with multiple Skill invokes). Each role with

--skip-questions

For ui-designer: append

--style-skill {package}

Template-Driven Analysis:

• Load role-specific template if exists (e.g.,

  templates/role-templates/system-architect-template.md

  )
• Inject template into agent prompt as required structure
• For system-architect: MUST include Data Model, State Machine, Error Handling, Observability, Configuration Model, Boundary Scenarios

Quality Validation:

• After analysis generation, invoke

  role-analysis-reviewer-agent

  to validate against template
• Check MUST have sections (blocking), SHOULD have sections (warning), quality checks (RFC keywords, valid diagrams)
• Output validation report with score and recommendations

TodoWrite: Attach N parallel sub-tasks, execute concurrently, collapse on completion.

Phase 3.5: UI Design Exploration

Ref: phases/03a-ui-design-exploration.md

Sequential phase after text-role analyses complete. Generates tangible UI artifacts (HTML prototypes, ASCII mockups, or API sketches) based on project type detection.

Input: all role analyses from Phase 3 (read-only), guidance-specification.md, style_skill (optional) Output:

html-prototypes/

ascii-mockups/

api-sketches/

Skip Conditions: project type is library/service, no ui-designer in selected_roles,

--yes

--ui-explore

Optional Review: If 2+ prototypes and not

--yes

Ref: ~/.ccw/workflows/brainstorm-visualize.md

TodoWrite: Attach single task with sub-tasks (Detection → Generate → Validate → Review → Feedback), collapse on completion.

Phase 4: Synthesis Integration

Ref: phases/04-synthesis.md

Six-phase synthesis: Discovery → File discovery → Cross-role analysis → User interaction → Spec generation (single agent, sequential) + Conditional review → Finalization. Role analysis files are read-only (never modified). Spec is the final synthesis artifact.

Input: session_id from Phase 2, all role analysis files from Phase 3 (read-only) Output: feature-specs/, feature-index.json, synthesis-changelog.md

TodoWrite: Attach synthesis sub-tasks, execute sequentially, collapse on completion.

Single Role Mode Execution (execution_mode = "single-role")

Phase 3: Single Role Analysis

Ref: phases/03-role-analysis.md

Execute role analysis for ONE specified role with optional interactive context gathering.

Input: role_name, --session, --update, --include-questions, --skip-questions, --style-skill Output: {role}/analysis*.md

TodoWrite: Attach 4 sub-tasks (Detection → Context → Agent → Validation), execute sequentially.

Phase Reference Documents (read on-demand when phase executes):

Core Rules

1. Start with Mode Detection: First action is Phase 1 (parse args + detect mode)
2. Interactive Routing: If mode cannot be determined from args, ASK user via AskUserQuestion
3. No Preliminary Analysis: Do not analyze topic before Phase 2 - artifacts handles all analysis
4. Parse Every Output: Extract selected_roles from workflow-session.json after Phase 2
5. Auto-Continue via TodoList: Check TodoList status to execute next pending phase automatically
6. Task Attachment Model: Skill and Task executes attach sub-tasks to current workflow
7. ⚠️ CRITICAL: DO NOT STOP: Continuous multi-phase workflow in auto mode. After executing all attached tasks, immediately collapse them and execute next phase
8. Parallel Execution: Auto mode Phase 3 attaches multiple agent tasks simultaneously for concurrent execution
9. Single Role Independence: Single role mode operates independently without requiring artifacts or synthesis

Input Processing

Parameter Parsing

// Parse from user input (argument string)
const args = parseArguments(user_input);

// Flags
const auto_yes = args.includes('--yes') || args.includes('-y');
const count = extractFlag(args, '--count', 3);  // default 3, max 9
const session_id = extractFlag(args, '--session', null);
const update_mode = args.includes('--update');
const include_questions = args.includes('--include-questions');
const skip_questions = args.includes('--skip-questions');
const style_skill = extractFlag(args, '--style-skill', null);

// Role detection
const VALID_ROLES = [
  'data-architect', 'product-manager', 'product-owner', 'scrum-master',
  'subject-matter-expert', 'system-architect', 'test-strategist',
  'ui-designer', 'ux-expert'
];
const first_arg = args[0]; // first non-flag argument
const is_role = VALID_ROLES.includes(first_arg);

// Mode detection
if (auto_yes) {
  execution_mode = 'auto';
  topic = extractTopic(args);
} else if (is_role) {
  execution_mode = 'single-role';
  role_name = first_arg;
} else {
  execution_mode = null;  // Ask user
  topic = extractTopic(args);
}

Usage Examples

# Auto mode - full pipeline
/brainstorm "Build real-time collaboration platform" --count 3
/brainstorm -y "GOAL: Build platform SCOPE: 100 users" --count 5
/brainstorm "Design payment system" --style-skill material-design

# Single role mode - individual analysis
/brainstorm system-architect --session WFS-xxx
/brainstorm ux-expert --include-questions
/brainstorm ui-designer --session WFS-xxx --update --style-skill material-design
/brainstorm product-manager --skip-questions

# Ambiguous - will ask interactively
/brainstorm --session WFS-xxx
/brainstorm

Data Flow

Phase 1 (Mode Routing):
  Input:  user arguments
  Output: execution_mode, parsed_params
          ↓
  ┌───────┴───────┐
  Auto            Single Role
  ↓               ↓
Phase 2:          Phase 3:
  Input:  topic, count, auto_yes
  Output: session_id ─────────────→ Input: role_name, session_id
          selected_roles[]                  skip/include questions
          guidance-specification.md         style_skill
          ↓                         Output: {role}/analysis*.md
Phase 3:
  Input:  selected_roles[], session_id
          guidance-specification.md
          style_skill (for ui-designer)
  Output: {role}/analysis*.md (N files, immutable after this point)
          ↓
Phase 3.5:
  Input:  all analysis*.md (read-only), guidance-specification.md, style_skill
  Output: html-prototypes/ OR ascii-mockups/ OR api-sketches/
          feature-index.json, review_decisions (optional)
          ↓
Phase 4:
  Input:  session_id, all analysis files (read-only)
  Output: feature-specs/F-{id}-{slug}.md
          feature-index.json
          synthesis-changelog.md

TodoWrite Pattern

Auto Mode Lifecycle

Initial → Phase 1 Mode Routing (completed)
       → Phase 2 Artifacts (in_progress)
           → 7 sub-tasks ATTACHED (Phase 0-5)
           → Execute sequentially
           → Sub-tasks COLLAPSED
       → Phase 3 Parallel Role Analysis (in_progress)
           → N role sub-tasks ATTACHED simultaneously (excluding ui-designer)
           → Execute concurrently
           → Sub-tasks COLLAPSED
       → Phase 3.5 UI Design Exploration (in_progress)
           → Sub-tasks: Detection → Generate → Validate → Review → Feedback
           → Execute sequentially
           → Sub-tasks COLLAPSED
       → Phase 4 Synthesis (in_progress)
           → 8 sub-tasks ATTACHED
           → Execute sequentially
           → Sub-tasks COLLAPSED
       → All completed

Single Role Mode Lifecycle

Initial → Phase 1 Mode Routing (completed)
       → Phase 3 Role Analysis (in_progress)
           → 4 sub-tasks ATTACHED (Detection → Context → Agent → Validation)
           → Execute sequentially
           → Sub-tasks COLLAPSED
       → Completed

Initial TodoWrite (Auto Mode)

[
  {"content": "Phase 1: Mode detection and parameter parsing", "status": "in_progress", "activeForm": "Detecting mode"},
  {"content": "Phase 2: Interactive Framework Generation", "status": "pending", "activeForm": "Generating framework"},
  {"content": "Phase 3: Parallel Role Analysis", "status": "pending", "activeForm": "Executing parallel analysis"},
  {"content": "Phase 3.5: UI Design Exploration", "status": "pending", "activeForm": "Generating UI prototypes"},
  {"content": "Phase 4: Synthesis Integration", "status": "pending", "activeForm": "Executing synthesis"}
]

Initial TodoWrite (Single Role Mode)

[
  {"content": "Phase 1: Mode detection and parameter parsing", "status": "in_progress", "activeForm": "Detecting mode"},
  {"content": "Phase 3: Single role analysis for {role_name}", "status": "pending", "activeForm": "Executing role analysis"}
]

Session Management

⚡ FIRST ACTION: Check

.workflow/active/

Multiple Sessions Support:

• Different Claude instances can have different brainstorming sessions
• If multiple sessions found, prompt user to select
• If single session found, use it
• If no session exists:
  • Auto mode: Create

    WFS-[topic-slug]

  • Single role mode: ERROR if no session (must run auto mode first)

Session Continuity: All phases share session state via

workflow-session.json

Available Roles

data-architect

product-manager

product-owner

scrum-master

subject-matter-expert

system-architect

test-strategist

ui-designer

ux-expert

Role Selection: Auto mode → handled by artifacts (Phase 2). Single role mode → user specifies directly.

Output Structure

.workflow/active/WFS-{topic}/
├── workflow-session.json              # Session metadata ONLY
├── .process/
│   └── context-package.json           # Phase 0 output (auto mode)
└── .brainstorming/
    ├── guidance-specification.md      # Framework (Phase 2, auto mode)
    ├── feature-index.json             # Feature index (Phase 4, auto mode, feature_mode)
    ├── synthesis-changelog.md         # Synthesis decisions audit trail (Phase 4, auto mode)
    ├── feature-specs/                 # Feature specs (Phase 4, auto mode, feature_mode)
    │   ├── F-001-{slug}.md
    │   └── F-00N-{slug}.md
    ├── specs/
    │   └── terminology-template.json  # Terminology schema
    ├── templates/
    │   └── role-templates/
    │       └── system-architect-template.md  # System architect analysis template
    ├── agents/
    │   └── role-analysis-reviewer-agent.md   # Role analysis validation agent
    ├── html-prototypes/               # Phase 3.5 (web/mobile/desktop-gui)
    │   ├── README.md                  # Prototype summary
    │   ├── feature-index.json         # File-to-feature mapping
    │   └── {feature-slug}.html        # Self-contained prototypes
    ├── ascii-mockups/                 # Phase 3.5 (cli projects)
    │   └── {feature-slug}.md
    ├── api-sketches/                  # Phase 3.5 (library/service projects)
    │   └── {feature-slug}.md
    ├── {role}/                        # Role analyses (IMMUTABLE after Phase 3)
    │   ├── {role}-context.md          # Interactive Q&A responses
    │   ├── analysis.md                # Main/index document
    │   ├── analysis-cross-cutting.md  # Cross-feature (feature_mode)
    │   └── analysis-F-{id}-{slug}.md  # Per-feature (feature_mode)
    └── synthesis-specification.md     # Integration (Phase 4, non-feature_mode only)

Error Handling

Context Overflow Protection:

• Per-role limits: < 3000 words main, < 2000 words sub-docs, max 5 sub-docs
• Synthesis protection: If total > 100KB, read only

  analysis.md

  (not sub-documents)
• Recovery: reduce scope (--count 2) → use --summary-only → manual synthesis

Coordinator Checklist

Pre-Phase Actions:

• Read Phase document via

  Ref:

  marker
• Verify prerequisites (session exists, required files present)
• Mark phase as

  in_progress

  in TodoWrite
• Attach sub-tasks if applicable

Post-Phase Actions:

• Validate phase outputs exist
• Collapse sub-tasks to phase summary
• Mark phase as

  completed

  in TodoWrite
• Auto-continue to next pending phase (auto mode)
• Report completion (single role mode)

Related Commands

Prerequisites:

• /workflow:session:start

  - Start a new workflow session (optional, brainstorm creates its own)

Follow-ups (after brainstorm completes):

• /workflow-plan --session {sessionId}

  - Generate implementation plan

• /workflow:brainstorm:synthesis --session {sessionId}

  - Run synthesis standalone (if skipped)

Reference Information

Template Source:

~/.ccw/workflows/cli-templates/planning-roles/

.claude/skills/style-{package-name}/