#Controlling the Output Format¶

If code consumes the output, the format is a contract — engineer it like one.

Techniques, Weakest → Strongest¶

1. Describe the format¶

1Answer with a comma-separated list of country names only.

Cheap, but the model may still add "Sure! Here you go:".

2. Show the exact shape¶

1Respond ONLY with JSON in exactly this shape:
2{"country": string, "capital": string, "population": number}
3No prose, no markdown fences.

3. Few-shot the format¶

Demonstrate 2–3 input→output pairs where the output is exactly the target format. The model copies structure extremely reliably from examples.

4. Prefill / response priming¶

Start the model's answer for it so it has no room to add preamble:

1... Output the JSON now.
2Assistant: {

5. Use the API's structured-output feature¶

Most providers offer JSON mode or schema-constrained / structured outputs that guarantee valid JSON matching a schema. When available, this is strictly better than prompting alone — use it (covered hands-on in Module 3).

Robust Parsing Still Matters¶

Even with the above, defensively parse:

1function safeJson(text) {
2  // Strip accidental \`\`\`json fences before parsing
3  const cleaned = text.replace(/^\`\`\`(json)?|\`\`\`$/g, "").trim();
4  return JSON.parse(cleaned);
5}

Common Format Failures & Fixes¶

Golden rule: The more your downstream code depends on the format, the stronger the technique you should use — and always parse defensively.