ClaudeMarketing/pipeline/CLAUDE.md

Marketing Content Pipeline

This project implements an AI-powered Social Media Content Automation System. Eight specialized agents research, generate, render, and distribute marketing content.

System Architecture

Eight agents running in sequence:

1. Trend Scout — trending content monitoring via Tavily
2. Marketing Research Agent — deep market research via Tavily
3. Script Writer — ad scripts from research output
4. Gemini Ad Designer — AI-generated image ads via NanoBanana MCP (Google Gemini)
5. Poster Ad Designer — museum-quality poster ads via Playwright HTML-to-PNG
6. Video Ad Producer — video ads via Remotion
7. Copywriter Agent — platform-specific copy
8. Distribution Agent — publish manifest creation (gate-protected)

Folder Structure

• assets/ — brand images, logos, product shots (mood board)
• knowledge/ — brand identity, platform guidelines, product/campaign info
• skills/ — all 8 agent skills (each has SKILL.md)
• outputs/ — generated content per campaign
• remotion-ad/ — Remotion video project with compositions

Knowledge Files (READ FIRST)

Every agent MUST read these before generating ANY content:

• _knowledge/brand_identity.md — tone, voice, personality, CTA patterns
• _knowledge/platform_guidelines.md — per-platform specs and formatting
• _knowledge/product_campaign.md — product details, features, campaign direction

NOTE: The _knowledge/ directory is dynamically written per-pipeline-run with the active app's content. The original knowledge/ directory contains the honeyDue defaults.

Available Tools

Tavily (Web Research)

Use @tavily/core npm package. Write a Node.js script to run searches:

import { tavily } from "@tavily/core";
const client = tavily({ apiKey: process.env.TAVILY_API_KEY });
const result = await client.search("query", { maxResults: 10 });

The TAVILY_API_KEY is available in the environment.

NanoBanana MCP (Image Generation)

Available as MCP tool: mcp__nanobanana__generate_image Uses Google Gemini to generate images. Call it with a detailed prompt describing the desired image. The GEMINI_API_KEY is configured in .mcp.json.

Playwright (HTML to PNG)

Use Playwright to render HTML/CSS layouts to pixel-perfect PNG screenshots:

import { chromium } from "playwright";
const browser = await chromium.launch();
const page = await browser.newPage();
await page.setViewportSize({ width: 1080, height: 1080 });
await page.setContent(htmlString);
await page.screenshot({ path: "output.png" });
await browser.close();

Remotion (Video Rendering)

Video project is in remotion-ad/. Compositions are defined in remotion-ad/src/. To render a video:

cd remotion-ad && npx remotion render src/index.ts CompositionId --output ../outputs/campaign/video/filename.mp4

You can modify or create new compositions in remotion-ad/src/ before rendering.

Pipeline Execution Order

trend-scout → research → script-writer → gemini-ad-designer → poster-ad-designer → video-producer → copywriter → distribution

Each agent reads its SKILL.md from skills/{agent-name}/SKILL.md and follows it exactly.

Output Convention

outputs/{task_name}_{YYYYMMDD}/
  ├── trend_report.json
  ├── research_results.json
  ├── research_brief.md
  ├── interactive_report.html
  ├── ads/
  │   ├── instagram_feed_*.png (1080x1080)
  │   ├── instagram_stories_*.png (1080x1920)
  │   ├── nextdoor_spotlight_*.png (1200x1200)
  │   ├── nextdoor_display_*.png (1200x628)
  │   └── ad_manifest.json
  ├── scripts/
  │   ├── scripts_all.json
  │   └── scripts_summary.md
  ├── video/
  │   ├── instagram_reel_*.mp4 (1080x1920)
  │   ├── tiktok_ad_*.mp4 (1080x1920)
  │   ├── nextdoor_video_*.mp4 (1080x1080)
  │   └── scene_plans.json
  ├── copy/
  │   ├── instagram_captions.json
  │   ├── tiktok_captions.json
  │   ├── nextdoor_posts.json
  │   └── copy_matrix.json
  └── Publish_manifest.md

File Naming Convention

{platform}_{format}_{hook_variant}_{dimensions}.{ext} Example: instagram_feed_hook_a_1080x1080.png

Safety Rules

• No live API posting without explicit user approval
• Distribution agent creates a publish manifest — NEVER auto-publishes
• All media files must exist locally before creating the manifest
• Always save outputs to the specified output directory