Skip to content
Academy
Marketing Academy · Field Work●AI in Marketing
All projects·From: No-Code Marketing Automation: n8n, Zapier & Make
MiniBuild the Asset· 30 minutes

Draft the LLM Node: Turning a Vague Prompt Into a Working Spec

RXBAR

Objective: Given a plain data-moving workflow (form submission to Slack), write a complete LLM node spec, task, input mapping, and structured output schema, that a no-code platform could actually run without producing inconsistent results.

You're the marketing ops lead at RXBAR, the Chicago-founded protein bar company acquired by Kellogg's for $600M. Your website's 'wholesale inquiry' form dumps raw text into a Slack channel, and reps skip long submissions.

Write the LLM node's task instruction, input mapping, and output schema so two different runs of the same submission produce the same classification, following the lesson's three-part prompt structure.

Before you start

What you'll need

Free path (everything below is enough to finish)

Google Sheets
FreeDraft and iterate the prompt text before pasting it into the platform

Free, easy to share with a second reviewer before it goes live

Zapier
FreemiumBuild and test the actual LLM node with structured output

Free tier supports a limited number of Zaps with AI actions, enough to build and test one workflow

No access? n8n's free self-hosted tier if Zapier's free-tier task limit is too tight

The process

2 steps

Step 01 of 02

Stating the exact task instead of a vague instruction

The lesson's three-part prompt structure requires stating the exact task ('classify into exactly one of these three categories'), not a vague instruction like 'analyze this.'

A first-draft prompt reads 'Look at this wholesale inquiry and tell us what to do.' Rewrite it as an exact task instruction.

Google Sheets— Draft the prompt text in a shared sheet before pasting it into the n8n or Zapier LLM node.

Procedure

  1. List the 3 categories reps actually route on: high-volume-retail, small-batch-inquiry, spam-or-irrelevant
  2. Write one sentence naming the task and the closed category list
  3. Add one line telling the model to also extract the requested case-pack quantity if present
  4. Read the instruction back, could a different person run it twice and get the same category both times?
Sample output
TASK INSTRUCTION (draft)
Classify this wholesale inquiry into exactly one of: high-volume-retail, small-batch-inquiry, spam-or-irrelevant. If a case-pack quantity is mentioned, extract it as a number.

Healthy

The instruction names a closed list of categories and a single extraction field.

Unhealthy

The instruction says 'tell us what to do,' leaving the category set undefined.

What this means

A closed category list is what makes two runs of the same input agree, an open-ended instruction invites a different answer every time.

So what do I do about it?

SymptomActionEffort
Reps see inconsistent tags on similar submissionsRewrite the prompt to name the exact closed category list30 min
YouYou can do this yourself, no engineering access required.

Step 02 of 02

Specifying a structured output format the next node can parse

Structured output modes constrain the model to return valid JSON matching a schema, guaranteeing the field names and types the next node expects, instead of a paragraph you have to regex apart.

The Slack-posting node downstream needs a category string and a numeric quantity field. Write the JSON schema the LLM node should be told to return.

Zapier— The LLM node's 'response format' or 'structured output' setting inside the Zap editor.

Procedure

  1. Define the schema: category as an enum of the 3 values, quantity as a number or null
  2. Paste the schema into the LLM node's structured output field
  3. Map the Slack message template to read {{category}} and {{quantity}} directly, no parsing step
  4. Send one test submission with no quantity mentioned, confirm quantity returns null, not an empty string
Sample output
{
  "category": "high-volume-retail" | "small-batch-inquiry" | "spam-or-irrelevant",
  "quantity": number | null
}

Healthy

Every test run returns valid JSON with exactly these two fields, no free text wrapper.

Unhealthy

The model sometimes returns a sentence like 'This looks like a retail inquiry for 500 units.'

What this means

A schema-less prompt makes the next node's parsing brittle, one unexpected sentence format breaks the whole chain downstream.

So what do I do about it?

SymptomActionEffort
Slack message shows raw unparsed model outputAdd the structured output schema to the LLM node before mapping downstream fields30 min
YouYou can do this yourself, no engineering access required.

Final deliverable

A written LLM node spec (task instruction + input mapping + JSON output schema) ready to paste into a no-code platform.

See a reference example
Sample output
Blue Bottle Coffee, wholesale inquiry LLM node spec

TASK: Classify this wholesale inquiry into exactly one of: high-volume-retail, small-batch-inquiry, spam-or-irrelevant. Extract case-pack quantity if mentioned.

INPUT MAPPING: {{form.message}} -> model input

OUTPUT SCHEMA:
{
  "category": "high-volume-retail" | "small-batch-inquiry" | "spam-or-irrelevant",
  "quantity": number | null
}

TEST RUN 1 (500-unit cafe order): {"category": "high-volume-retail", "quantity": 500}
TEST RUN 2 (same input, re-run): {"category": "high-volume-retail", "quantity": 500}

Success criteria

You're done when you can:

  • Task instruction names a closed, finite category list
  • Output schema is valid JSON with typed fields the next node can map directly
  • Same test input produces the same classification on repeated runs