Design Swarm

Design Swarm

Orchestrate 10 specialized design agents — each embodying a different skill/lens — to run holistic audits, ideation sessions, and implementations in parallel.

Orchestration Flow

When invoked, follow these steps exactly:

1. Parse Arguments

Extract from the invocation:

• Modes: One or more of audit, ideate, implement (default: audit if none specified)
• Target: The thing to evaluate — one of:
  • Screenshot path (.png, .jpg, .webp)
  • URL (http:// or https://)
  • File path (code file)
  • Text brief (quoted description of an interface)

2. Detect Target Type

3. Load Orchestration Context

Read these files:

• modules/orchestrator.md — synthesis logic, conflict resolution, final report template
• references/scoring-rubric.md — universal 1-10 scale
• references/agent-weights.md — default weights for composite scoring

4. Create Team

TeamCreate(team_name: "design-swarm", description: "Holistic design audit/ideation/implementation")

5. Execute Phases

For each requested mode, in order (audit → ideate → implement):

Phase: Audit

1. Determine participating agents (see orchestrator.md Phase Gating):

   • Screenshot-only: skip baseline-ui
   • Code-only: all 10
   • URL: all 10 (snapshot first, then inspect code if accessible)
   • Brief: all 10 (theoretical assessment)

2. Create tasks via TaskCreate — one per participating agent

3. Spawn agents in parallel via Task tool. For each agent:

   a. Read the agent template: modules/agents/<agent-name>.md b. Read the source knowledge file(s) listed in the template's "Source Knowledge" section c. Construct the agent prompt by combining:

   • The full agent template (structure + evaluation criteria)
   • The full source knowledge content
   • The target context (screenshot/URL snapshot/code/brief)
   • Mode: "audit"
   • The scoring rubric content
   • Instructions to mark task complete and send report via SendMessage to "team-lead"

   Task(
     name: "<agent-name>",
     team_name: "design-swarm",
     subagent_type: "general-purpose",
     run_in_background: true,
     prompt: [constructed prompt above]
   )
   

4. Wait for all agents to complete (they'll send reports via SendMessage)

5. Synthesize per orchestrator.md:

   • Collect all reports
   • Deduplicate findings
   • Aggregate by severity
   • Compute weighted composite score
   • Identify and resolve conflicts
   • Generate the final audit report

6. Present the synthesized audit report to the user

Phase: Ideate

1. Create tasks for each agent (blocked by audit synthesis if audit was run)

2. Spawn agents with the same process as audit, but:

   • Mode: "ideate"
   • Include the synthesized audit findings in each agent's prompt
   • Each agent proposes improvements through their specific lens

3. Synthesize ideation proposals:

   • Merge overlapping proposals
   • Categorize by impact/effort matrix
   • Produce prioritized improvement roadmap

4. Present the improvement roadmap to the user

Phase: Implement

1. Spawn a single general-purpose agent with:

   • The full synthesized audit report
   • The prioritized improvement roadmap
   • References to all relevant source skill files
   • Instructions to implement changes in priority order (critical → serious → moderate → minor)

2. Present implementation results to the user

6. Shutdown

After all phases complete:

1. Send shutdown_request to all active agents
2. Call TeamDelete to clean up

Agent Reference

Usage Examples

/design-swarm audit ~/Desktop/screenshot.png
/design-swarm audit ideate https://example.com
/design-swarm audit ideate implement ./src/components/Settings.tsx
/design-swarm ideate "dark mode settings panel with toggle, theme preview, and schedule"
/design-swarm audit implement ./src/pages/Dashboard.tsx

Important Notes

• Runtime knowledge loading: Agent templates are compact. Source knowledge is read fresh from canonical skill files at spawn time. This prevents knowledge duplication and ensures agents always use the latest skill content.
• Parallel execution: All agents within a phase run simultaneously. Phases run sequentially when chained.
• Conflict resolution: The craft-philosophy agent (weight 1.5) serves as tiebreaker. See orchestrator.md for the full protocol.
• Partial modes: Any combination of modes is valid. ideate without prior audit uses the target directly (no findings context). implement without ideate applies audit findings directly.