Pre-fill Responses

What is Prefill?

Prefill (or "putting words in Claude's mouth") is a technique where you start the assistant's response with specific text. The model then continues from where you left off.

As Anthropic explains:

> "Parsing XML tags is nice, but maybe you want a structured JSON output to make sure it's JSON serializable. You could just add that Claude needs to begin his output with a certain format."

How Prefill Works

API Structure

const response = await anthropic.messages.create({
  model: "claude-sonnet-4-20250514",
  max_tokens: 1024,
  messages: [
    {
      role: "user",
      content: "Analyze this product review..."
    },
    {
      role: "assistant",
      content: "{"  // Prefill starts here
    }
  ]
});

// Response continues: "sentiment": "positive", ...}

The Model Continues

Common Prefill Patterns

JSON Object

messages: [
  { role: "user", content: prompt },
  { role: "assistant", content: "{" }
]
// Model completes: {"field": "value", ...}

JSON Array

messages: [
  { role: "user", content: "List the top 5 issues..." },
  { role: "assistant", content: "[" }
]
// Model completes: ["issue1", "issue2", ...]

XML Tags

messages: [
  { role: "user", content: prompt },
  { role: "assistant", content: "<final_verdict>" }
]
// Model completes: <fault>Vehicle B</fault>...</final_verdict>

Code

messages: [
  { role: "user", content: "Write a function to calculate..." },
  { role: "assistant", content: "

]

// Model completes: total):\n return sum(items)...\n

Why Prefill is Powerful

1. Format Guarantee

Without prefill, the model might add preamble:

Here's my analysis of the document:

{
  "sentiment": "positive",
  ...
}

With prefill ({), you get clean JSON:

{
  "sentiment": "positive",
  ...
}

2. Reduced Tokens

No need for "Output only JSON..." instructions when you prefill with {.

3. Role Guidance

Prefill can set the tone:

{ role: "assistant", content: "As a senior engineer reviewing this code, " }

Combining with Output Format

Full Example

const systemPrompt = `
You analyze product reviews and return structured data.

Output format:
{
  "sentiment": "positive" | "negative" | "mixed",
  "confidence": 0.0-1.0,
  "key_points": ["point1", "point2"]
}
`;

const userPrompt = `
Review to analyze:
"This laptop is amazing! Fast, lightweight, but the battery
could be better. Still highly recommend."
`;

const response = await anthropic.messages.create({
  model: "claude-sonnet-4-20250514",
  system: systemPrompt,
  messages: [
    { role: "user", content: userPrompt },
    { role: "assistant", content: "{" }
  ]
});

// Clean JSON output guaranteed
const result = JSON.parse("{" + response.content[0].text);

Prefill for Different Formats

Structured Text

{ role: "assistant", content: "## Analysis\n\n" }
// Model continues with markdown content

Code with Language

{ role: "assistant", content: "

// Model writes TypeScript code

### Specific Structure

{

role: "assistant",

content: '{"status": "complete", "analysis": {'

}

// Forces nested JSON structure

## When to Use Prefill

| Situation | Use Prefill? |
|-----------|--------------|
| Need clean JSON | Yes, start with `{` |
| Need XML output | Yes, start with `<tag>` |
| Conversational response | No |
| Format already specified | Optional |
| Code generation | Yes, for language hint |

## Prefill Gotchas

### 1. Don't Over-Prefill

// Too much - constrains the model

{ content: '{"sentiment": "positive"' }

// Just right - guides format

{ content: "{" }

### 2. Match Your Parser

// Prefill

{ content: "{" }

// Your parser must prepend the prefill

const json = JSON.parse("{" + response.content[0].text);

### 3. Handle Incomplete Responses

// If max_tokens is hit, JSON may be incomplete

try {

const data = JSON.parse("{" + response);

} catch {

// Handle truncation

}

## Real-World Application

For the car accident analysis:

const systemPrompt = You analyze Swedish car accident forms...;

const userPrompt =

const response = await anthropic.messages.create({

model: "claude-sonnet-4-20250514",

system: systemPrompt,

messages: [

{ role: "user", content: userPrompt },

{ role: "assistant", content: "" }

]

});

// Parse the verdict

const verdictContent = "" + response.content[0].text;

```

Key Takeaways

• Prefill guides format - Start the response to control structure
• Use for JSON/XML - Guarantees clean, parseable output
• Reduces instructions - Less "output only..." text needed
• Combine with format spec - Prefill + examples = reliable output
• Handle in parser - Prepend the prefill to complete the response

Next Lesson

In our final lesson, we'll explore Extended Thinking - how to use Claude's reasoning capabilities to improve and debug prompts.