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

Signal	Type	Notes
.png, .jpg, .webp, .gif	screenshot	Read the image directly
http://, https://	url	Use browser tools to snapshot the page
Existing file path	code	Read the file(s)
Everything else	brief	Treat as text description

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"