Shape a GPT-6.1 brief, then paste into Create

Split the jobs: GPT-6.1 Sol turns a messy request into a visual brief; GPT Image 2.5 renders on Create. Do not treat them as one button. Product-line context: GPT-6.1 Sol vs GPT Image.

Default sample: ecommerce product hero (wireless earbuds). For posters or portraits, swap only Subject and Scene.

One-line flow

1. Open GPT-6.1 Sol in ChatGPT Work / Codex / API
2. Paste the system prompt below plus your draft ask
3. Receive a five-line brief (one line per layer + Avoid)
4. Open Create, pick GPT Image 2.5 Flare, paste the brief, set ratio and quality
5. After 2–3 trials, switch to Sunburst; lock SKU with references when needed

Copyable: five-line brief template

Edit only the text after each colon:

Subject: matte-black true wireless earbuds, case half-open, one bud upright outside
Scene: seamless light-gray tabletop, clean PDP hero, no clutter props
Light: soft side light, thin specular on the case, short clean shadow
Frame: 3:4 vertical, subject centered slightly low, margin for a price tag later
Avoid: no readable brand type in-frame, no watermark, no second product

This matches the site’s four-layer habit—see prompt structure.

Copyable: system prompt for GPT-6.1

Use as system / developer text (or the first message):

You are a visual-brief editor, not an image model.
Compress the user request into exactly 5 lines starting with Subject:/Scene:/Light:/Frame:/Avoid:
(If the user writes Chinese, use 主体:/场景:/光线:/构图:/禁止: instead.)
Add only material, light, and framing facts; drop marketing fluff.
Never name Flare/Sunburst, credits, API model ids, pixel sizes, or whether to upload references.
Output only those 5 lines—no commentary.

Example user message:

Need a clean PDP hero for new ANC earbuds; show a bit of the case; detail-page ready.

Expected shape: five lines only, ready to paste into the Create prompt box.

Paste into Create: checklist

Open Create and tick in order:

Step	Action
1	Model: GPT Image 2.5 Flare (explore first)
2	Ratio from the Frame line (sample uses 3:4 or the closest on-site option)
3	Quality: default / auto first; raise later for finals
4	Paste the five lines only—no chat context
5	Generate 2–3 times; change one layer per run (e.g. Light only)
6	Lock direction → Sunburst; see Flare vs Sunburst
7	Series heroes that must hold shape → attach a product reference, then refine

Three anti-patterns

1. Pasting the full chat log — noise like “try again” drifts the frame.
2. Asking GPT-6.1 to “generate the image / base64” — it is not the on-site image engine; finals must go through Create billing and receipts.
3. Writing “use Sunburst 4K” inside the brief — pick tier and resolution in Create controls; stuffing them into prose fights the subject line.

Switching categories: change two lines

• Poster: Subject → hero figure or title object; Scene → night city / stage; Frame → leave room for type.
• Portrait lock: Subject → face + wardrobe detail; Scene → simplify; must attach a face reference.
• Keep the same system prompt; only swap the user draft and the five lines.

Next step

1. Run one earbud loop: GPT-6.1 → five-line brief → Create Flare → Sunburst.
2. Still mixing product lines? Read GPT-6.1 vs GPT Image.
3. Prompts still mushy? Rebuild the skeleton with four-layer structure, then let GPT-6.1 compress.