How to fix a language model refusing to answer with system prompts in n8n?

The fix is to explicitly send the “system prompt” as part of the model’s messages array, not as a separate field, because most modern language models (including n8n’s OpenAI and OpenAI‑Compatible nodes) ignore any “system” field outside of the messages array. So in n8n you must pass your system instructions as a message with role = system. If you're feeding your messages through expressions or other nodes, make sure the JSON you send to the LLM node includes a proper system role message and that you're not overwriting it later in the workflow.

 

Why the model refuses system prompts in n8n (and how to fix it)

 

Modern LLMs (OpenAI, OpenAI‑compatible, Cohere, etc.) accept instructions only through a messages array. n8n’s “OpenAI” node and most “Chat Model” nodes do not magically force system instructions if you put them somewhere else (for example, in a header, config field, or a variable called systemPrompt). If the model doesn’t see a JSON message with role: system, it simply behaves like the system instruction never existed.

In production workflows, this usually breaks when:

• You build the message array in a previous Function node and accidentally drop the system message.
• You pass only the user message into an LLM node, thinking n8n will merge your system prompt automatically – it won’t.
• Some LLMs (like OpenAI’s newer models) ignore deprecated “system” fields unless it is placed exactly in the messages list.
• You modify messages with an expression like {{$json}} and overwrite the entire structure.

 

How to fix it reliably

 

You must explicitly send the system instruction as a message object inside the messages array. The model will obey it only if it’s formatted exactly like this.

Here is a safe pattern using a Function node before the LLM node:

 

// This Function node builds clean LLM messages
return [
  {
    json: {
      messages: [
        {
          role: "system",
          content: "You are a helpful assistant. Follow these rules strictly."
        },
        {
          role: "user",
          content: "Explain how to fix the issue."
        }
      ]
    }
  }
];

 

Then, in your OpenAI / OpenAI-Chat / OpenAI-Compatible node, set:

• Resource: Chat
• Operation: Create Chat Completion
• Messages: Use expression like {{$json["messages"]}}

This forces the node to send the system message exactly as the model expects.

 

Extra production‑grade notes

 

• If you build messages dynamically (from a database, webhook, or long workflow), always log the final JSON before it hits the model. The “lost system prompt” is the most common cause of unexpected model behavior.
• Don’t put system instructions in n8n’s “Prompt” field of the old OpenAI node if you’re using chat-completion mode. That field becomes a user message and the system instruction is ignored.
• Some models enforce strict roles. If you send something like role: "instruction" or role: "assistant" with system content, the system message will be ignored.
• If your system prompt is extremely long, consider storing it in an environment variable or reading it from a file to avoid mistakes when editing workflows.
• After fixing, run a test execution and check the “RAW API request” section to confirm the message is being sent with role: system. This is the best way to debug LLM behavior in n8n.

 

Bottom line

 

The model isn’t refusing system prompts — it never received them. In n8n, the only reliable fix is to include the system prompt as a normal { role: "system", content: "..." } object inside the messages array. Once you construct the messages cleanly and pass them to the LLM node without overwriting them, the model will follow the instructions consistently.