chain-llm-pattern

Chain LLM Pattern for n8n

Multi-step LLM chains outperform single-shot prompts on any task that combines extraction + reasoning. This skill encodes the production pattern.

When to use a chain vs a single prompt

Single prompt works	Chain is better
"Summarize this email"	"Extract entities, then categorize by urgency, then decide routing"
"Translate this to English"	"Detect language, translate, then extract structured fields"
"Is this spam? yes/no"	"Score spam probability from email, phone, IP, content separately, then combine"

Rule of thumb: if the task has ≥2 distinct reasoning steps OR the final decision depends on intermediate structured data, use a chain.

The pattern (3-stage default)

Input → [Extract] → [Analyze/Classify] → [Score/Decide] → Output

Each stage is its own LLM node with its own prompt. Between stages, use Set or Code nodes to transform and validate.

Stage 1: Extract (schema-bound)

Use Information Extractor node (LangChain). NOT a generic AI Agent or raw HTTP call.

Why: Information Extractor binds output to a JSON schema. It parses, retries on invalid JSON, and fails loudly — instead of silently returning prose you then regex.

Define schema explicitly:

{
  "type": "object",
  "properties": {
    "customer_name": { "type": "string" },
    "product_mentioned": { "type": "string" },
    "sentiment": { "enum": ["positive", "neutral", "negative"] },
    "urgency_score": { "type": "number", "minimum": 0, "maximum": 10 }
  },
  "required": ["customer_name", "sentiment"]
}

System prompt for this stage: short, one job. "Extract the fields defined in the schema from the transcript. If a field is absent, omit it. Do not infer or guess."

Stage 2: Analyze (reason over extracted data)

Use Basic LLM Chain with the extracted JSON from Stage 1 as input.

This stage reasons: categorize, cluster, identify patterns, detect issues. The input is structured (from Stage 1) so the model isn't juggling parsing + reasoning simultaneously.