I’ve been experimenting with using Ollama for content generation in n8n, and found that chaining multiple focused AI calls produces much better results than a single big prompt.
Here’s the approach — a 4-stage pipeline where each node refines the previous output:
The Pipeline
Stage 1: Research → Generate key points about the topic (temperature: 0.7)
Stage 2: Outline → Structure those points into a logical flow (temperature: 0.6)
Stage 3: Draft → Write the full content (temperature: 0.8)
Stage 4: Edit → Polish grammar, flow, and clarity (temperature: 0.3)
Why Multi-Stage Works Better
When you ask an LLM to “write a blog post about X” in one shot, it tends to:
- Rush through important points
- Produce shallow content
- Miss logical structure
By breaking it into stages, each call has a focused job and builds on quality input from the previous stage.
Quick Setup
Each stage is just an HTTP Request node hitting Ollama’s API:
POST http://localhost:11434/api/generate
The key is the prompt engineering at each stage. Here’s the research stage as an example:
{
"model": "llama3:8b",
"prompt": "You are a research assistant. Research the following topic and provide 5-7 key points that should be covered in a comprehensive blog post. Include specific facts and statistics where relevant.\n\nTopic: {{ $json.topic }}",
"stream": false,
"options": { "temperature": 0.7, "num_predict": 1024 }
}
Each subsequent stage references {{ $json.response }} from the previous HTTP Request output.
Model Recommendations
| Model | VRAM | Best For |
|---|---|---|
llama3:8b |
~5GB | Good all-around, fast |
mistral |
~4GB | Concise, fast |
llama3:70b |
~40GB | Highest quality |
Tips
- Set
"stream": false— n8n needs the complete response, not chunks - Use different temperatures per stage — creative for drafting, conservative for editing
- Add
"num_predict": 4096for the draft stage to avoid truncation - If Ollama is in Docker, use
http://host.docker.internal:11434instead of localhost
Has anyone else built multi-stage AI pipelines in n8n? I’d love to hear what patterns you’ve found useful!