coefficient-portfolio

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

Agent 2: Creative

Agent 3: Editor-in-Chief

Agent 4: Analyst (Planned)

Key Features

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

Key Decisions

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.