Reliable JSON Output Prompt AI Prompt

Design prompts and parsing strategies to get reliable, parseable JSON from LLMs every time.

Unreliable JSON is one of the most common LLM integration failure modes — the model adds markdown fences, explanatory text, trailing commas, or truncates the output mid-JSON.

1. Prompt instructions for reliable JSON:

   Instruction 1 — Format command:
   'Return ONLY a JSON object. Do not include any explanation, markdown formatting, or code blocks.'

   Instruction 2 — Schema specification:
   'The JSON must match this exact schema: {{json_schema}}'
   Include a JSON Schema definition or a clear field-by-field description with types.

   Instruction 3 — Null handling:
   'If a field cannot be determined from the input, set it to null. Do not omit fields.'

   Instruction 4 — No truncation:
   'Return the complete JSON object. Never truncate. If the output would be very long, summarize field values rather than cutting off.'

   Instruction 5 — Validation example:
   Append a valid example at the end of the prompt: 'Your output should look like this: {{example_json}}'

2. Engineering safeguards (client-side):

   Safeguard 1 — JSON extraction from messy output:
   Even with good prompts, models sometimes add preamble. Use regex to extract JSON:
   ```python
   import re, json
   def extract_json(text):
       match = re.search(r'\{[\s\S]*\}', text)
       if match:
           return json.loads(match.group())
       raise ValueError('No JSON found in output')
   ```

   Safeguard 2 — Schema validation:
   After extraction, validate against the expected schema using jsonschema or Pydantic.

   Safeguard 3 — Retry with correction:
   If JSON parsing fails: re-call the model with: 'Your previous response was not valid JSON. The error was: {{error}}. Return only the corrected JSON object.'

   Safeguard 4 — Structured output APIs:
   Use model APIs that enforce JSON output natively (OpenAI response_format, Anthropic tool use, Instructor library).

3. Model-specific tips:
   - Add 'Your response:' followed by '{' at the end of the prompt to prime the model to start with JSON
   - For long JSON objects: request the model output one section at a time and merge

Return: the reliable JSON prompt template, extraction code, schema validation code, retry logic, and a test harness that measures JSON parse success rate across 100 calls.