1. The Golden Rule of JSON Prompting: Never assume the model knows what you want. Bad prompt: ``` "Return a JSON with user info" ``` Good prompt: ``` Return a JSON object with exactly these fields: { "name": "string - full name", "email": "string - valid email address",

2. Schema First, Always Define your schema before writing prompts. Use this template: ``` json { "field_name": "type - description with constraints", "status": "enum - one of: pending|completed|failed", "confidence": "number - float between 0.0 and 1.0", "metadata":

3. Validation Patterns That Actually Work Enum constraints: "priority": "enum - high|medium|low" Range limits: "score": "number - integer 1-10" Format specs: "date": "string - ISO format YYYY-MM-DD" Required vs optional: Mark clearly with "required" or "optional" The model

4. Error Handling in JSON Prompts Always include error handling instructions: ``` If you cannot extract the requested information: { "error": "string - describe what's missing", "partial_data": "object - any data you could extract", "confidence": 0.0 } ``` This prevents

5. Multi-step Agent JSON Structure For complex agents, use nested action objects: ``` { "reasoning": "string - explain your thinking", "action": { "type": "enum - search|call|write|analyze", "parameters": "object - action-specific data" }, "next_step": "string -

6. Common JSON Prompting Mistakes: ❌ Using vague field names (info, data, result) ❌ No type specifications ❌ Missing example outputs ❌ No fallback for missing data ❌ Ignoring token limits in large schemas ✅ Specific names, explicit types, clear examples, error handling,

7. Testing Your JSON Prompts Run these tests before deploying: 1. Happy path - perfect input 2. Missing data - incomplete information 3. Ambiguous data - unclear inputs 4. Edge cases - unusual but valid scenarios 5. Malformed requests - broken input data Your prompt should

8. Performance Optimization Smaller JSON schemas = faster responses = lower costs. Instead of: ``` { "user_first_name": "string", "user_last_name": "string", "user_email_address": "string" } ``` Use: ``` { "name": "string - first and last name", "email": "string" }

9. Advanced: Conditional JSON Schemas For agents that need different outputs based on context: ``` If task_type is "research": Return: {"findings": [...], "sources": [...]} If task_type is "analysis": Return: {"summary": "...", "recommendations": [...]} If task_type is

10. The One Trick That Changed Everything Add this to every JSON prompt: ``` Before outputting JSON, think through your response step by step. Then output ONLY valid JSON, no additional text. ``` This reduces malformed JSON by 80%. The model self-corrects before committing.

Real-World Example: Customer Service Agent { "intent": "enum - question|complaint|request|compliment", "urgency": "enum - low|medium|high|critical", "category": "string - product|billing|technical|other", "sentiment": "number - -1.0 to 1.0 scale", "response": "string -

Reliable agents start with bulletproof JSON prompts. Specify everything, test everything, validate everything. The 5 minutes you spend perfecting your schema saves hours of debugging broken agents.

10x your prompting skills with my prompt engineering guide → Mini-course → Free resources → Tips & tricks Grab it while it's free ↓ https://godofprompt.ai/prompt-...

That's a wrap: I hope you've found this thread helpful. Follow me @godofprompt for more. Like/Repost the quote below if you can: