design-agent

name

design-agent

description

Use when the user wants to produce a complete design specification bundle — visual direction, design brief, mood board, brand rules, image generation prompts, accessibility audit, and asset acceptance checklist — for a campaign, product, or brand. Triggers when the user says "design a campaign", "create design specs", "design agent", "/design-agent", or describes a visual direction / brand identity / asset brief task. Invokes the AgentSuite Design agent via MCP.

Design Agent (AgentSuite)

This skill invokes the Design agent from the AgentSuite MCP server. It produces 9 design specification artifacts and 8 production-ready asset brief templates in 30–120 seconds and pauses for human approval before promoting to long-lived storage.

When to use

User wants any of:

• Visual direction document
• Design brief for a campaign
• Mood board specification
• Brand rules extraction from existing materials
• Image generation prompts for AI art tools (Midjourney, Stable Diffusion, DALL·E)
• Asset acceptance checklist
• Accessibility audit template (WCAG AA)
• Ready-to-fill brief templates for banner ads, social graphics, landing heroes, email headers, deck slides, print flyers, video thumbnails, or icon sets

When NOT to use

• One-off copy or text tasks — use the Founder agent or write directly
• Technical product specification — that's the Product agent (v0.3+)
• Marketing campaign execution — that's the Marketing agent (v0.5+)
• Brand system / voice guide from scratch — run the Founder agent first, then the Design agent reads from _kernel/

Steps

1. Confirm AgentSuite MCP is configured. Call agentsuite_list_agents. If "design" is not in enabled, tell the user to set AGENTSUITE_ENABLED_AGENTS=founder,design in their MCP env config (paste the snippet from ~/.claude/skills/design-agent/mcp-snippet.json).

2. Gather inputs from the user. Ask:

   • Target audience (one sentence)
   • Campaign goal (one sentence)
   • Output channel: web / social / email / print / video / deck / other
   • Project slug (lowercase, hyphenated) — for _kernel/ promotion
   • Path to brand materials directory (optional — PDFs, images, markdown docs)

3. Invoke design_run. Pass the gathered fields plus agent_name="design", role_domain="design-ops". Wait for the JSON-RPC response (30–120 s; do not poll).

4. Show the user the result. Display the primary_path (always visual-direction.md) and the list of open_questions. Tell them to read visual-direction.md and qa_report.md.

5. Wait for approval signal. When the user says "approved" or "ship it", call design_approve with the same run_id and project_slug.

6. Confirm promotion. Echo the promoted_paths from the ApprovalResult so the user sees what landed in _kernel/.

Cost expectation

A typical run costs $0.30 – $1.50 against Claude Sonnet or GPT-4o (9 LLM calls for spec artifacts + 1 consistency check + 1 QA score). Hard cap is $5.00 per run. If HardCapExceeded is raised, reduce input size or raise AGENTSUITE_COST_CAP_USD.

Failure modes

• NoProviderConfigured — tell user to set ANTHROPIC_API_KEY or OPENAI_API_KEY.
• ConsistencyCheckFailed — open consistency_report.json and explain the mismatch between artifacts.
• extract stage produced invalid JSON — re-run, usually transient.
• QA score below threshold — requires_revision=true in result; open qa_report.md and revision-instructions.md for specific changes.

After approval

The promoted artifacts in _kernel/<slug>/ can be fed directly into any image generation tool, design tool (Figma, Canva), or LLM session as grounding context. The brief-template-library/ folder contains 8 ready-to-fill brief templates for specific asset types.