AI Content Production Pipeline

Chapter 17: AI Content Production Pipeline

Content creation is one of the fastest AI automation wins. A well-designed n8n workflow can collect trending topics, filter and rank them, generate full articles, create cover images, and publish to multiple platforms — all automatically. This chapter builds the complete pipeline end-to-end, with detailed API cost analysis.

17.1 Pipeline Overview

Five stages, each implemented as a distinct section of the n8n workflow (or separate sub-workflows for easier debugging):

1. Step 1 — Trend Collection: RSS Feed + Google Trends API
2. Step 2 — AI Filtering + Dedup: LLM relevance scoring + fingerprint deduplication
3. Step 3 — Content Generation: Claude/GPT with structured prompt templates
4. Step 4 — Image Generation: DALL-E 3 or Stable Diffusion API
5. Step 5 — Scheduled Publishing: Platform-specific APIs for multi-channel distribution

17.2 Step 1: RSS + Google Trends Collection

n8n's native RSS Read node handles any RSS/Atom feed without XML parsing code. Configure multiple feed URLs — n8n merges all entries automatically. For Google Trends, use the HTTP Request node to call SerpAPI's trends endpoint.

GET https://serpapi.com/search.json
  ?engine=google_trends
  &q=n8n+automation,AI+workflow
  &date=now+7-d
  &geo=US
  &api_key={{ $credentials.serpApiKey }}

# Extract rising queries from related_queries.rising
# Use these as content angle suggestions

17.3 Step 2: AI Filtering and Deduplication

Use GPT-4o-mini to score each RSS item for relevance (0–10) at roughly $0.0001 per item. A Filter node keeps only items scoring 7 or above. A Code node then fingerprints titles (first 15 chars, normalized) to remove near-duplicates, then sorts by score and keeps the top 5.

// Code node: deduplication by title fingerprint
const items = $input.all();
const seen = [];
const deduped = [];

for (const item of items) {
  const fingerprint = item.json.title.toLowerCase()
    .replace(/[^a-z0-9]/g, '').slice(0, 15);
  if (!seen.includes(fingerprint)) {
    seen.push(fingerprint);
    deduped.push(item);
  }
}

deduped.sort((a, b) => b.json.aiScore - a.json.aiScore);
return deduped.slice(0, 5);

17.4 Step 3: Content Generation

Claude 3.5 Sonnet outperforms GPT-4o on Chinese long-form writing and follows style instructions more reliably. Use a structured prompt that specifies: author persona, target platform, article structure, word count, and style constraints.

Cost tip: For budget-conscious setups, use Claude 3 Haiku for first drafts ($0.25/M tokens), human-review the top picks, then publish. Quality is lower but acceptable for high-volume informational content.

17.5 Step 4: Image Generation

// DALL-E 3 API call
{
  "method": "POST",
  "url": "https://api.openai.com/v1/images/generations",
  "body": {
    "model": "dall-e-3",
    "prompt": "{{ $json.imagePrompt }}",
    "size": "1792x1024",
    "quality": "standard"
  }
}

// DALL-E 3 standard 1792x1024: $0.080/image
// Self-hosted Stable Diffusion: ~$0.002/image on a GPU VPS

Before calling DALL-E, use an LLM node to translate the Chinese article title into an English image prompt. DALL-E 3 interprets English prompts far more accurately than Chinese.

17.6 Step 5: Multi-Platform Publishing

n8n has native nodes for WordPress and Notion. Use HTTP Request for platforms with public APIs.

Editorial safeguard: Never auto-publish raw AI output to public platforms. Write generated content to a Notion database with status "pending review." After human approval, a second n8n workflow (triggered by Notion status change) handles the actual publishing. This preserves quality control without sacrificing automation.

17.7 Cost Analysis

5 articles per day, ~1,200 words each. Per-article API costs:

5 articles/day = ~$0.57/day or ~$17/month. Swap DALL-E for a self-hosted Stable Diffusion instance and the monthly cost drops under $5.