How to Fix a Language Model Refusing to Answer with System Prompts in n8n

Open your AI Agent or Basic LLM Chain node and check where the system prompt is configured. In the AI Agent node, the system prompt goes in the System Message field under the top-level options, not in the user's input. In the Basic LLM Chain node, use the System Message input. A common mistake is pasting the system prompt into the Prompt (user message) field, which means the model treats it as a user request rather than a system instruction. When the system prompt is in the user field, the model may interpret statements like 'You are a customer service agent who only discusses product X' as a user asking it to roleplay, which can trigger refusals for certain topics. Move the system prompt to the dedicated System Message field to ensure correct role assignment.

Models refuse requests when the system prompt contains contradictory instructions. For example, a system prompt that says 'Never discuss competitor products' followed by 'Always provide helpful comparisons when asked' creates a conflict that the model resolves by refusing. Review your system prompt line by line and remove or reconcile any contradictions. Pay special attention to negation-heavy instructions. Instead of telling the model what NOT to do, tell it what to DO. Replace 'Do not refuse to answer questions about pricing' with 'Always answer pricing questions directly using the provided data.' Positive instructions are less likely to trigger safety filters than negative ones that inadvertently describe disallowed behaviors.

1// BAD system prompt (contradictory, negation-heavy):
2// "You are a helpful assistant. Never refuse to answer. 
3//  Do not discuss harmful topics. Always be comprehensive. 
4//  Keep responses short."
5
6// GOOD system prompt (clear, positive, non-contradictory):
7// "You are a customer support agent for Acme Corp.
8//  Answer questions about Acme products using the knowledge base provided.
9//  For topics outside Acme products, say: 'I can only help with Acme product questions.'
10//  Keep responses under 200 words."

When your use case involves topics that might trigger safety filters, such as medical information or legal advice with disclaimers, add explicit permission statements to the system prompt. These statements tell the model that it is authorized to discuss these topics within defined boundaries. For example, if you are building a medical FAQ bot, your system prompt should include something like 'You are authorized to provide general health information from the attached knowledge base. Always include the disclaimer that this is not medical advice.' This gives the model explicit context that the topic is approved within the conversation. Without this, models default to their built-in safety behavior which refuses medical, legal, and financial topics.

In multi-turn conversations managed by the AI Agent node, incorrect message ordering can cause refusals. The LLM expects messages in the order: system, user, assistant, user, assistant. If your workflow sends messages out of order, or sends two user messages in a row without an assistant response between them, the model can behave unpredictably. When using Memory sub-nodes like Window Buffer Memory or Postgres Chat Memory, verify that the stored conversation history maintains correct alternation between user and assistant roles. Add a Code node before the AI Agent that validates and repairs message ordering if needed.

1// Code node: Validate and repair message ordering
2const items = $input.all();
3
4for (const item of items) {
5  const messages = item.json.messages || [];
6  const repaired = [];
7  let lastRole = 'system';
8
9  for (const msg of messages) {
10    if (msg.role === 'system') {
11      repaired.push(msg);
12      lastRole = 'system';
13    } else if (msg.role === 'user' && lastRole !== 'user') {
14      repaired.push(msg);
15      lastRole = 'user';
16    } else if (msg.role === 'assistant' && lastRole === 'user') {
17      repaired.push(msg);
18      lastRole = 'assistant';
19    }
20    // Skip duplicate consecutive roles
21  }
22
23  item.json.messages = repaired;
24}
25
26return items;

Different LLM providers handle system prompts differently in n8n. OpenAI Chat Model accepts system prompts as a separate message role. Anthropic Claude uses a dedicated system parameter outside the messages array. Google Gemini uses system_instruction. When you switch models in your n8n workflow, the system prompt may not transfer correctly between providers. If you are using the AI Agent node, n8n handles this translation automatically. But if you are using HTTP Request nodes to call LLM APIs directly, you must format the request body according to each provider's spec. Verify by checking the raw request in the execution data.

1// OpenAI format (HTTP Request body):
2{
3  "model": "gpt-4o",
4  "messages": [
5    { "role": "system", "content": "You are a helpful assistant." },
6    { "role": "user", "content": "{{ $json.userMessage }}" }
7  ]
8}
9
10// Anthropic format (HTTP Request body):
11{
12  "model": "claude-3-5-sonnet-20241022",
13  "system": "You are a helpful assistant.",
14  "messages": [
15    { "role": "user", "content": "{{ $json.userMessage }}" }
16  ],
17  "max_tokens": 1024
18}

Use n8n's manual execution feature to test prompt changes quickly. Run the workflow manually, check the LLM output, adjust the system prompt, and run again. Use data pinning to lock the input data so you can test different prompt configurations against the same user message. Pin the data on the node before the LLM by clicking the pin icon in the output panel. This freezes that node's output so downstream nodes always receive the same input during testing. This approach lets you iterate on the system prompt without needing to trigger the full workflow from the start each time. Track which prompt versions work and which cause refusals.