#4. Structured Output Enforcer

The Problem

Your team built a Q&A agent that answers user questions. Downstream systems expect the agent's response to be a JSON object with three keys — answer, confidence, and sources — but the current prompt just says "answer the question," so the model returns free-text paragraphs. Every consumer that tries to JSON.parse() the response crashes. The model is capable of producing JSON; the issue is the system prompt never asks for it. Your job is to fix only the system prompt so every response is a valid JSON object with the required keys.

Examples

Example 1

User input: What is the capital of France?

Current (bad) output: The capital of France is Paris. It's been the capital since the 10th century and is known for the Eiffel Tower, the Louvre, and its café culture.

Expected (good) output:

{
  "answer": "The capital of France is Paris.",
  "confidence": 0.99,
  "sources": ["general knowledge"]
}

Example 2

User input: Who wrote the novel 1984?

Current (bad) output: George Orwell wrote 1984, published in 1949. It's a dystopian novel about totalitarianism.

Expected (good) output:

{
  "answer": "George Orwell wrote the novel 1984, published in 1949.",
  "confidence": 0.98,
  "sources": ["general knowledge"]
}

Example 3

User input: What is the half-life of Carbon-14?

Current (bad) output: The half-life of Carbon-14 is approximately 5,730 years. This property makes it useful for radiocarbon dating.

Expected (good) output:

{
  "answer": "The half-life of Carbon-14 is approximately 5,730 years.",
  "confidence": 0.95,
  "sources": ["general knowledge", "nuclear physics"]
}

Your Task

Update the system prompt (and only that) so the agent:

• Always returns a valid JSON object as its entire response.
• Includes exactly three keys: answer (string), confidence (number between 0 and 1), and sources (array of strings).
• Produces meaningful values — not empty strings or placeholders.
• Never wraps the JSON in markdown code fences or adds text outside the JSON.

Do not add tools, change model parameters, or add post-processing code.

Evaluation

Submissions are checked for the following:

• Valid JSON output: The agent's full response is parseable JSON with no extra text.
• Correct keys present: The JSON object contains answer, confidence, and sources.
• Values are meaningful: Each field has a substantive, relevant value — not empty or boilerplate.