Turn Rough Ideas Into JSON Image Briefs
You get a clean, fully specified JSON brief plus a one-line summary you can paste into any image generator. Splitting the idea into labelled fields forces every detail to be stated, so the model has less room to guess.
The prompt
You are an art director who writes precise briefs for AI image generators.
Task: Turn my rough idea into a structured JSON image brief that an image model can follow without guessing.
Rough idea: {rough_idea}
Era and mood: {era_and_mood}
Camera and lighting: {camera_and_lighting}
Steps:
1. Split the idea into four top-level keys: subject, photography, background and avoid.
2. Inside each key, add short nested fields such as clothing, lens, props or light source. Fill each with one concrete visual detail instead of a vague adjective.
3. Apply the era, mood, camera and lighting details consistently across every key.
4. Add a fixed_details list naming anything that must stay exactly as described, such as colours, shapes or printed text.
5. Check the brief for contradictions and fix them before you answer.
Output: Valid JSON only, with no commentary. After the JSON, add one final line beginning "Prompt:" that sums up the brief in a single sentence I can paste into an image tool.
Change the parts in curly braces. Use it when your image results keep drifting from what you pictured, or when you need many images in the same consistent style.
What happened when we ran it
We sent
You are an art director who writes precise briefs for AI image generators. Task: Turn my rough idea into a structured JSON image brief that an image model can follow without guessing. Rough idea: A small bakery owner stands behind her counter holding a tray of cinnamon buns. The shop is cosy and a little cluttered, with handwritten price cards and a chalkboard menu. I want it to look like an authentic photo for our website homepage, not a stock image. Era and mood: Present day, warm, friendly and unfussy, early morning before opening time Camera and lighting: 35mm lens, shallow depth of field, soft window light from the left, slight film grain, shot at eye level Steps: 1. Split the idea into four top-level keys: subject, photography, background and avoid. 2. Inside each key, add short nested fields such as clothing, lens, props or light source. Fill each with one concrete visual detail instead of a vague adjective. 3. Apply the era, mood, camera and lighting details consistently across every key. 4. Add a fixed_details list naming anything that must stay exactly as described, such as colours, shapes or printed text. 5. Check the brief for contradictions and fix them before you answer. Output: Valid JSON only, with no commentary. After the JSON, add one final line beginning "Prompt:" that sums up the brief in a single sentence I can paste into an image tool.
The model replied
Run on Claude Sonnet 5.5, 29 Oct 2026. Automatic check: 4 out of 5.
Mentioned for
Gemini, Nano Banana Pro, ChatGPT, Claude
Where it came from
Found on GitHub, shared by ZeroLu: ZeroLu/awesome-nanobanana-pro.
This is our own version of the idea, written and tested by us. The original may differ.