Orchestrate a multi-agent design workflow to create or iterate on a design that meets project requirements.

Arguments

•
$1: Design name (required) - Slug for the design subdirectory (e.g., "user-auth", "payment-flow")
- •Creates/uses directory: memory-bank/planning/{name}/
- •If not provided, ask user for the design name or list existing designs
•
$2: Mode or focus area (optional)
- •new - Force creation of a new design from scratch (ignores existing design.md)
- •Any other text - Focus iteration on this specific area (e.g., "ux", "architecture", "types")
- •If not provided and design.md exists - Run general iteration to improve design
- •If not provided and no design.md - Run full new design workflow

Design Directory

All artifacts for a design are stored in memory-bank/planning/{name}/:

code

memory-bank/planning/{name}/
├── design.md                     # Current design document
├── design-draft.md               # Working draft (new design workflow)
├── research-*.md                 # Research agent outputs
├── review-*.md                   # Review outputs (new design)
└── iteration-review-*.md         # Iteration review outputs

Mode Selection

First, validate the design name and determine which mode to run:

code

if $1 is empty:
    → Use Glob to find existing designs: memory-bank/planning/*/design.md
    → If found, list them and ask user to select or provide a name
    → If none found, ask user for a new design name
    → Store selected/provided name as DESIGN_NAME

DESIGN_NAME = $1
DESIGN_DIR = memory-bank/planning/{DESIGN_NAME}

if $2 == "new":
    → Run NEW DESIGN WORKFLOW
else if file_exists("{DESIGN_DIR}/design.md"):
    → Run ITERATION WORKFLOW (with optional focus: $2)
else:
    → Run NEW DESIGN WORKFLOW

Project Type Detection

Before running agents, detect project characteristics:

•TypeScript: Check if tsconfig.json exists OR grep memory-bank/project-brief.md for "TypeScript"
•Expo: Check if app.json contains "expo" key OR grep memory-bank/project-brief.md for "Expo"
•Design System: Check if memory-bank/DESIGN-AGENTS.md contains design-system-file: and the referenced file exists

Store detection results for agent selection.

Agent Selection

Build the list of agents to use:

Universal agents (always include):

•ai-resources:design:requirements-validation-agent
•ai-resources:design:software-engineering-agent
•ai-resources:design:simplicity-advocate-agent
•ai-resources:design:testing-strategy-agent
•ai-resources:design:ux-usability-agent

Conditional agents (based on detection):

•If TypeScript detected: add ai-resources:design:typescript-type-safety-agent
•If Expo detected: add ai-resources:design:expo-best-practices-agent
•If Design System configured: add ai-resources:design:design-system-agent

Project-specific agents:

•Read memory-bank/DESIGN-AGENTS.md if it exists
•Add any agents listed under "## Research Agents" (for Phase 2)
•Add any agents listed under "## Review Agents" (for Phase 4)
•These are project-local agents in .claude/agents/

NEW DESIGN WORKFLOW

Use this workflow when creating a design from scratch.

Phase 1: Context Gathering

Read these files to understand the project context:

•memory-bank/project-brief.md - Project overview and tech stack
•
Find requirements file:
- •Check memory-bank/DESIGN-AGENTS.md for requirements-file: config
- •If not found, glob for memory-bank/*requirements*.md or memory-bank/*REQUIREMENTS*.md
•If design system configured, read the design system spec file
•Read any relevant source files referenced in requirements

If requirements file cannot be found, use AskUserQuestion to ask for its location.

Phase 2: Research Phase (Parallel Agents)

If memory-bank/DESIGN-AGENTS.md lists any Research Agents, launch them ALL IN PARALLEL:

For each research agent, use this prompt template:

code

Research [agent's domain] for this project.

Project context: [summary from Phase 1]
Requirements: [key requirements]

Write your findings to `{DESIGN_DIR}/research-{agent-name}.md`

Include:
- Executive Summary
- Key Findings
- Recommendations
- Tradeoffs and Alternatives
- References

Wait for all research agents to complete before proceeding.

If no research agents are configured, skip to Phase 3.

Phase 3: Design Synthesis

Read all research outputs from {DESIGN_DIR}/research-*.md, then create an initial design document.

Create {DESIGN_DIR}/design-draft.md with:

•Metadata - Version, date, status
•Summary - Executive overview
•Requirements Addressed - How each requirement is addressed
•Architecture - Component structure, data flow
•Data Model - Types, schemas, state shape
•Service Layer - Service interfaces and responsibilities
•UI Components - Screen hierarchy, component breakdown (if applicable)
•Algorithms - Key algorithms and their rationale (if applicable)
•User Flows - Key user journeys
•Testing Strategy - Approach for each layer
•Design Decisions - Key decisions with rationale
•Open Questions - Unresolved issues

Phase 4: Design Review (Parallel Agents)

Launch ALL review agents IN PARALLEL using the Task tool with multiple tool calls in a single message:

For each review agent (universal + conditional + project-specific), use this prompt:

code

Review the design at `{DESIGN_DIR}/design-draft.md` against [agent's specialty].

Also read:
- Requirements file: [path]
- Project brief: memory-bank/project-brief.md
[If design-system-agent: - Design system: [path]]

Write your review to `{DESIGN_DIR}/review-{agent-name}.md`

Follow your standard output format.

Wait for all reviews to complete.

Phase 5: Design Refinement

Read all review outputs from {DESIGN_DIR}/review-*.md.

•Synthesize feedback - Combine findings from all agents
•Resolve conflicts - If agents disagree, document tradeoffs and make a decision
•Update design-draft.md - Address Critical and Major issues
•Document decisions - Add entries to Design Decisions section

If significant changes were made (Critical issues addressed), optionally re-run Phase 4 reviews. Maximum 2 review cycles total.

Phase 6: Final Design Document

Copy the refined design-draft.md to {DESIGN_DIR}/design.md.

Ensure it includes:

•Updated metadata (version, date, status: "Draft")
•Change Log with initial entry

Phase 7: Present to User

Present the design to the user:

•Summarize key design decisions (3-5 bullet points)
•Highlight major tradeoffs made
•List any open questions needing user input
•Ask for approval or feedback

ITERATION WORKFLOW

Use this workflow when {DESIGN_DIR}/design.md exists and you want to improve it.

Phase I-1: Load Context

Read these files:

•{DESIGN_DIR}/design.md - The existing design
•Requirements file (same logic as new design workflow)
•Any existing research files in {DESIGN_DIR}/research-*.md

Note the focus area from $2 if provided.

Phase I-2: Targeted Review (Parallel Agents)

Select agents based on focus area. If no focus provided, use ALL agents.

Focus area mapping:

•"ux", "ui", "usability", "interface" → ux-usability-agent, design-system-agent (if configured)
•"architecture", "code", "engineering" → software-engineering-agent, typescript-type-safety-agent, expo-best-practices-agent, testing-strategy-agent, simplicity-advocate-agent
•"requirements", "features" → requirements-validation-agent, simplicity-advocate-agent
•"type", "typescript", "types" → typescript-type-safety-agent
•"expo", "mobile" → expo-best-practices-agent
•"test", "testing" → testing-strategy-agent
•"simple", "complexity" → simplicity-advocate-agent
•No focus or "all" → ALL applicable agents

Launch selected agents IN PARALLEL with this prompt:

code

Review the existing design at `{DESIGN_DIR}/design.md`.

[If focus provided]: Focus specifically on: {focus_area}
[If no focus]: Provide a comprehensive review.

Your task:
1. Identify issues, gaps, or weaknesses
2. Suggest specific improvements with rationale
3. Note what's done well
4. Prioritize findings: Critical > Major > Minor > Suggestion

Write your review to `{DESIGN_DIR}/iteration-review-{agent-name}.md`

Format:
# {Agent Name} Review - Iteration

## Summary
[1-2 sentence overall assessment]

## Critical Issues
[Issues that would cause implementation failure]

## Major Issues
[Significant problems that should be fixed]

## Minor Issues
[Small improvements worth considering]

## Suggestions
[Nice-to-have improvements]

## Questions for Clarification
[Anything unclear needing input]

## What's Working Well
[Positive aspects]

Wait for all reviews to complete.

Phase I-3: Synthesize Feedback

Read all {DESIGN_DIR}/iteration-review-*.md files.

•Aggregate issues - Combine similar issues from multiple agents
•Prioritize - Order by severity and number of agents flagging it
•Identify conflicts - Note where agents disagree
•Determine scope - What can be addressed in this iteration?

If needed, use AskUserQuestion to:

•Resolve conflicting recommendations
•Clarify requirements ambiguities
•Get approval for significant changes
•Narrow scope if too many issues

Phase I-4: Apply Improvements

Update {DESIGN_DIR}/design.md:

•Critical issues - Must fix
•Major issues - Fix if straightforward
•Minor issues - Fix if quick
•Suggestions - Apply judgment

When making changes:

•Update relevant sections
•Add/update Design Decisions entries explaining changes
•Increment version in metadata
•Add entry to Change Log

Phase I-5: Validate Changes (Optional)

If significant changes were made, optionally re-run affected review agents. Limit to 1 validation pass.

Phase I-6: Present Changes to User

Summarize the iteration:

markdown

## Design Iteration Summary

### Changes Made

- [List of changes with brief rationale]

### Issues Addressed

- [Which issues were fixed]

### Issues Deferred

- [What was not addressed and why]

### New Questions

- [Any new questions that arose]

### Recommendation

[Next steps: another iteration, move to implementation, or specific area needs attention]

Ask user if they want to:

•Accept the changes
•Run another iteration (with optional focus)
•Revert changes
•Move to implementation

Agent Coordination Rules

•Parallel execution: Always launch independent agents in a SINGLE message with multiple Task tool calls
•Sequential dependencies: Wait for research before synthesis, synthesis before review
•Communication via files: Agents write to {DESIGN_DIR}/ directory
•Context efficiency: Agents read only files they need
•Max iterations: Run review cycle at most 2 times per workflow invocation
•Focus respect: When user provides focus, limit scope to relevant agents

Directory Structure

code

memory-bank/planning/{DESIGN_NAME}/
├── design.md                     # Current design document
├── design-draft.md               # Working draft (new design workflow)
├── research-*.md                 # Research agent outputs
├── review-*.md                   # Review outputs (new design)
└── iteration-review-*.md         # Iteration review outputs

Error Handling

•Missing design name: List existing designs or ask user to provide a name
•Invalid design name: Must be a valid slug (lowercase, alphanumeric, hyphens)
•Missing requirements file: Ask user for location via AskUserQuestion
•Missing DESIGN-AGENTS.md: Proceed with only universal + auto-detected agents
•Agent failure: Log error, continue with other agents, report at end
•No design.md for iteration: Switch to NEW DESIGN mode automatically
•Missing design directory: Create memory-bank/planning/{DESIGN_NAME}/ if it doesn't exist