AI Ad Creative System
Multi-agent system that researches Meta ad best practices from YouTube creators, then generates production-ready B2B ad creative with dual-version output (UI platform + AI model prompts).
What It Does
Built for a B2B SaaS startup targeting mid-market finance teams. Four specialized Claude agents collaborate to research ad creative patterns, generate campaign briefs, produce ad assets, and enforce quality standards. The system created 3 approved campaigns with 18 ad creatives each, all grounded in 160 research insights extracted from expert YouTube content.
Architecture
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Agent 1 │────▶│ Agent 2 │────▶│ Agent 3 │────▶│ Agent 4 │
│ Researcher │ │ Creative │ │ Editor │ │ Analyst │
│ │ │ │ │ │ │ (planned) │
│ YouTube → │ │ Brief-first │ │ 34-point │ │ Performance │
│ Insights → │ │ workflow → │ │ review → │ │ feedback │
│ Best │ │ Dual-version │ │ Approval │ │ loop │
│ Practices │ │ assets │ │ gate │ │ │
└──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘
│ │ │
▼ ▼ ▼
Knowledge Base Sub-Agents Review Report
(160 insights, (Video, Static, (34-point score,
38 best Infographic ≥31/34 to
practices) specialists) approve)
Agent Specifications
Agent 1: Researcher
- Role: YouTube content research specialist for B2B Meta ads
- Sources: 4 expert creators (Dara Denney, Denis Shatalin, Alex Cooper, Mr. Paid Social)
- Method: Dual-analysis (Claude for transcripts via Firecrawl, Gemini for visual analysis)
- Output: Living knowledge base with categorized insights (Hook, Visual, Copy, CTA, Format)
- Results: 160 insights extracted, 38 synthesized best practices, 3 videos analyzed
- Tools: YouTube MCP, Firecrawl MCP, Google Sheets MCP
Agent 2: Creative
- Role: Research-integrated creative director
- Workflow: Brief-first with user approval gates before asset generation
- Sub-agents: 3 specialists (Video Script, Static Ad, Infographic/Motion) run in parallel
- Output: Dual-version system:
- V1: Formatted for ad platform UI (adcreative.ai)
- V2: AI model prompts (GPT Image, Veo 3, Gemini)
- Key feature: Reads ALL 160 insights + 38 best practices before every brief, maps benefits to specific tactics with citations
Agent 3: Editor-in-Chief
- Role: Quality control and brand compliance gatekeeper
- Framework: 6-phase review (Quick scan → Best practices → Messaging → Brand → Production → Connector → Research alignment)
- Scoring: 34-point system across 6 sections, ≥31/34 (90%) required for approval
- Output: Structured review report with specific, actionable feedback
Agent 4: Analyst (Planned)
- Role: Performance monitoring and optimization feedback loop
- Would: Feed performance data back to Agent 1 to refine best practices
Key Features
- Brief-first workflow: User approves creative brief BEFORE any asset generation (prevents wasted work)
- Research grounding: Every creative decision cites specific insights (e.g., “BP_001: Statistics-Based Messaging”)
- Dual-version output: Every asset in 2 formats — platform UI workflows + AI model prompts
- Parallel sub-agents: 3 asset specialists run concurrently (45 min → 15-20 min)
- Knowledge base: Living document that grows with each video analyzed
- Messaging compliance: Finance-specific messaging guide enforced at every stage
Results
| Metric |
Value |
| Videos analyzed |
3 (primary) + 2 creator catalogs |
| Total insights extracted |
160 |
| Best practices synthesized |
38 |
| Campaigns created |
3 |
| Approval scores |
100%, 100%, 93% |
| Ad creatives per campaign |
18 (3 concepts × 3 sizes × 2 versions) |
| Review pass rate |
100% (all 3 approved) |
Stack
- Orchestrator: Claude Sonnet 4.5
- Sub-agents: Spawned via Task tool for specialized asset generation
- Video analysis: Claude (transcripts) + Gemini (visuals) via CCR routing
- Data: Google Sheets (structured) + Markdown knowledge base (narrative)
- MCP servers: YouTube, Firecrawl, Google Sheets, Browserbase, Google Docs
- AI tools: adcreative.ai (V1), GPT Image/Veo 3 (V2)
Key Decisions
- Brief-first over asset-first: Early sessions showed users rejected 60%+ of assets. Adding a brief approval gate before generation eliminated rework.
- Dual-version output: Not all teams have access to the same tools. V1 (platform UI) works for marketing teams; V2 (AI prompts) works for technical teams.
- YouTube as research source: Expert creators compress months of testing into 15-minute videos. Extracting their patterns via transcript + visual analysis is faster than running tests.
- 34-point scoring over subjective review: Quantified review criteria eliminated disagreements about “good enough” — either it passes 31/34 or it doesn’t.
- Living knowledge base over static rules: The best practices file grows with each video analyzed. Agent 2 reads ALL of it before every brief, so new research automatically improves output.
Visual Direction
For designer: 4-agent horizontal flow diagram with the knowledge base as a shared resource underneath. Show the brief-first approval gate between Agent 2’s brief and asset generation. Include the 3 parallel sub-agents branching from Agent 2. Highlight the 34-point scoring system in Agent 3’s box.