| name | json-prompt |
| description | Generate structured JSON prompts for any AI task - image generation, video generation, LLM structured outputs, product photography, UI mockups, commercial ads, and batch/programmatic generation. Use when user asks to create a JSON prompt, image prompt, video prompt, structured prompt, data extraction schema, batch template, or says "json prompt", "structured output", "extract data as JSON", "product photo prompt", "UI mockup prompt", "ad creative prompt", "batch generate". Covers visual AI platforms and LLM structured output patterns. |
| argument-hint | <image|video|llm|product|ui|ad|batch> <description> |
| allowed-tools | Read, Write, WebSearch, WebFetch |
JSON Prompt Generator
Generate structured JSON prompts across 7 domains. JSON prompting uses structured data instead of free-form text to achieve higher precision, consistency, and reproducibility from AI systems.
Step 1: Identify the Domain
| # | Domain | Trigger Phrases | Reference File |
|---|
| 1 | Image Generation | "image prompt", "generate a photo", "create a scene" | references/image-generation.md |
| 2 | Video Generation | "video prompt", "create a clip", "animate this" | references/video-generation.md |
| 3 | LLM Structured Output | "extract data", "structured response", "API output" | references/llm-structured-output.md |
| 4 | Product Photography | "product photo", "e-commerce shot", "packshot" | references/product-photography.md |
| 5 | UI/UX Mockups | "app mockup", "UI design", "screen layout" | references/ui-mockups.md |
| 6 | Commercial Advertising | "ad creative", "banner", "social media ad" | references/commercial-ads.md |
| 7 | Batch/Programmatic | "batch generate", "mass customize", "template with variables" | references/batch-programmatic.md |
If unclear, ask the user. Multiple domains can be combined (e.g., batch + product photography).
Step 2: Read the Reference File
Read the reference file for the identified domain. It contains the full JSON schema, examples, and domain-specific guidance. For batch requests, also read the primary domain's reference.
Step 3: For Visual Domains (1, 2, 4, 5, 6) — Select Target Platform
Read references/platform-optimization.md for platform-specific guidance. Key defaults:
- Images: Nano Banana Pro (best JSON parsing)
- Video: Veo 3 (best structured prompt support)
- LLM: Claude with structured outputs, or GPT-4o with
response_format
Step 4: Build the JSON Prompt
Use the schema from the reference file. Core principles:
- Separate subjects into distinct objects to prevent concept bleeding
- Use natural language descriptions within JSON values, not keyword tags
- Remove fields that don't apply (keep it lean)
- For multi-character scenes, always separate subjects into individual objects
Step 5: Validate and Present
- Validate JSON syntax — no missing commas, brackets, or quotes
- Check domain rules:
- Visual: consistent palette, no concept bleeding, no empty fields
- LLM: schema types declared, required fields marked, enums constrained
- Batch: every
{{variable}} has a matching key in the batch array
- Test readability — a human should understand the intent at a glance
Output format:
- Brief creative brief (1-2 sentences) explaining the intent
- Complete JSON prompt in a code block
- Platform-specific notes if relevant
- Suggestions for variations or next steps
- For batch: show the template + 2-3 example substitutions
Always present a single, complete JSON prompt — don't split across messages. Default to photorealistic for visual domains and "strict": true for LLM domains unless otherwise specified.
