How to Fix Broken Variable Interpolation in System Prompts in n8n

Every field in n8n that shows a small expression icon (the fx symbol) supports expressions. When you type {{ $json.name }} in a field, n8n evaluates this at runtime and replaces it with the actual value from the incoming data. The evaluation happens just before the node executes, using data from the most recent upstream node. Critical rules: expressions must be wrapped in {{ }}. The dollar sign variables $json, $input, and $('NodeName') are n8n-specific. If the expression cannot be resolved, n8n injects the string 'undefined' without throwing an error. This silent failure is the most common cause of broken interpolation because the workflow continues to run but the prompt contains the literal text 'undefined' instead of the expected value.

1// Correct syntax examples:
2{{ $json.userName }}              // Field from previous node
3{{ $('Webhook').first().json.body.message }}  // Field from named node
4{{ $input.first().json.email }}   // Explicit input reference
5{{ $execution.id }}               // Execution metadata
6
7// WRONG syntax:
8{ $json.userName }                // Missing outer braces
9{{ json.userName }}               // Missing $ prefix
10{{ $json['user name'] }}          // Correct for keys with spaces
11{{ $json.user.name }}             // Correct for nested objects

A common mistake is referencing a node that is downstream or on a different branch of the workflow. Expressions can only reference nodes that have already executed in the current run and are connected upstream via the data flow. If you reference $('DatabaseQuery') but the Database Query node is on a parallel branch that has not completed, the expression resolves to undefined. To fix this, ensure the referenced node is directly upstream of the LLM node. If data comes from a parallel branch, add a Merge node before the LLM node that combines both branches. Use the Merge node in Combine mode with Merge by Position to ensure all data is available before the prompt is built.

If your system prompt contains literal curly braces, for example a JSON template like '{"name": "value"}' that you want the LLM to follow as a format example, n8n's expression parser will try to evaluate everything inside {{ }}. This breaks the prompt. There are two solutions. First, if the curly braces are NOT inside {{ }}, they are treated as literal text in most n8n versions since v1.0. Only double curly braces {{ trigger expression evaluation. However, if you need to show the literal text {{ in your prompt, you must build the prompt in a Code node where you have full control over the string. Second, use a Code node to construct the entire prompt as a JavaScript string and output it as a field, then reference that field in the LLM node's system prompt.

1// Code node: Build system prompt with literal curly braces
2const userName = $input.first().json.userName || 'User';
3const outputFormat = '{"response": "your answer", "confidence": 0.95}';
4
5const systemPrompt = `You are a helpful assistant for ${userName}.
6
7Always respond in this exact JSON format:
8${outputFormat}
9
10Do not include any text outside the JSON object.`;
11
12return [{
13  json: {
14    systemPrompt: systemPrompt
15  }
16}];

When a variable might not exist in the incoming data, your expression should include a fallback. Without a fallback, the prompt receives the string 'undefined' which confuses the LLM. Use JavaScript's logical OR operator or nullish coalescing inside the expression to provide defaults. For example, {{ $json.userName || 'valued customer' }} returns 'valued customer' when userName is undefined, null, or empty. For more complex fallback logic, use a Code node that validates all required fields and assigns defaults before the data reaches the LLM node.

1// In expression fields, use OR for fallbacks:
2{{ $json.userName || 'there' }}
3{{ $json.language || 'English' }}
4{{ $json.context || 'No additional context provided.' }}
5
6// Code node for comprehensive validation:
7const item = $input.first().json;
8
9const validated = {
10  userName: item.userName || item.name || 'valued customer',
11  language: item.language || item.locale || 'English',
12  context: item.context || '',
13  maxLength: item.maxLength || 500,
14  tone: item.tone || 'professional'
15};
16
17return [{ json: validated }];

n8n's expression editor shows a live preview of the resolved value. Use this to verify that your expressions work before running the workflow. Click the fx icon next to any field to open the editor. On the left, browse available nodes and their output fields. On the right, see the expression and its resolved value. If the preview shows 'undefined' or an error, the expression is broken. Common issues visible in the preview include: the referenced node showing no data (it has not been executed in the current test), the field path being incorrect (typo in the property name), or the data type being wrong (referencing an array index that does not exist). Fix the expression until the preview shows the expected value, then close the editor.

For system prompts with multiple dynamic sections, conditional content, and formatted text, a Code node is more reliable and maintainable than inline expressions. Build the entire prompt as a JavaScript template literal in the Code node, referencing all necessary input data. This approach lets you use conditionals, loops, and string manipulation that are impossible with inline expressions. Output the completed prompt as a single field, then reference it in the LLM node's system prompt with a simple {{ $json.systemPrompt }} expression. This pattern separates prompt construction logic from the LLM configuration.

1// Code node: Complex System Prompt Builder
2const data = $input.first().json;
3
4const sections = [];
5
6// Role section
7sections.push(`You are a ${data.role || 'helpful assistant'} for ${data.company || 'our company'}.`);
8
9// Knowledge section (conditional)
10if (data.knowledgeBase) {
11  sections.push(`## Knowledge Base\nUse the following information to answer questions:\n${data.knowledgeBase}`);
12}
13
14// Rules section
15const rules = data.rules || ['Be concise', 'Be accurate'];
16sections.push('## Rules');
17rules.forEach((rule, i) => {
18  sections.push(`${i + 1}. ${rule}`);
19});
20
21// Output format section
22if (data.outputFormat === 'json') {
23  sections.push('## Output Format\nRespond ONLY with valid JSON. No text outside the JSON object.');
24}
25
26return [{
27  json: {
28    systemPrompt: sections.join('\n\n')
29  }
30}];