How do I stop LLMs from adding conversational text around a JSON output?

Prevent LLMs from adding extra text around JSON through specific prompting techniques. Use explicit instructions: "Output ONLY valid JSON with no additional text, explanations, or markdown formatting." Request "raw JSON" and specify "do not include backticks or code blocks." Enable JSON mode in APIs that support it (OpenAI, Anthropic) which constrains output to valid JSON only. Use structured output features when available for guaranteed JSON conformance. Place JSON schema in system prompt to reinforce format expectations. Add negative examples showing what NOT to do. Use stop sequences strategically though they may truncate valid JSON. Parse responses defensively: extract JSON between first { and last } if needed. Fine-tuned models follow formatting better than base models. Validate output with our JSON Validator at jsonconsole.com/json-editor before processing. For critical applications, use function calling or structured outputs rather than prompting. Clear, explicit instructions combined with validation create robust JSON extraction pipelines despite LLM tendency toward natural language.