skills/prompt-gpt-image/SKILL.md

---
name: prompt-gpt-image
description: Use when writing or improving a prompt for a Pika run that uses GPT Image — `create_image` or `edit_image` with model gpt-image-2.
---

# GPT Image prompts

Start from `prompt-image`; this adds what GPT Image rewards.

## In Pika

- Model id `gpt-image-2` in `create_image`. Fields besides `prompt`: `aspect_ratio`, `quality`, `background` (transparent for cut-outs and logos), `output_format`, and `image_urls` for edits and references. In `edit_image` the prompt goes in `instructions`, with `reference_urls`, `resolution` and `quality`.
- Put size, quality and transparency in their fields, not the prompt.

## Structure

- Lead with the result and its use: "A product photo for a store listing", "A poster for a jazz night".
- Describe what is visible: materials, textures, light, color, medium. Ask for "a real photograph" or "photorealistic" when that's the goal.
- For people: framing ("full body, feet visible"), gaze, pose, what they hold.
- For a long request, labeled sections read best: Scene, Subject, Details, Text, Constraints.

## Photoreal look

Describe it like a photographer: framing ("medium close-up at eye level"), lens ("50 mm, shallow depth of field"), film or grain, natural texture (pores, wear, creases), and "no heavy retouching" when it should look candid.

## Text in the image

- Exact words in quotes, with position and lettering: `the headline "OPEN LATE" in bold condensed white capitals across the top`.
- Spell unusual names letter by letter if they come back wrong.
- Ask for no other text. Check the spelling in the result; for small or dense text, try a higher `quality`.

## Edits and references

- Say "change only …", then list what must stay: identity, pose, geometry, label, lighting, colors, camera angle, background.
- Number the inputs and give each a role: "Image 1 is the person, image 2 is the jacket; dress the person in the jacket."
- One change per run; feed the result into the next edit and repeat what must be preserved.

## Examples

> A photorealistic candid photo of an elderly fisherman mending a net on a small wooden boat at dawn. Weathered skin with visible texture, a faded blue wool sweater. Medium close-up at eye level, 50 mm lens, soft low sun from the right, light film grain, no heavy retouching.

> Change only the background of image 1 to a sunlit white studio sweep. Keep the sneaker's shape, colors, laces and logo exactly, and keep the camera angle and shadows under the shoe.

> A minimal flat logo for a coffee cart named "Little Bean": a single coffee bean with a leaf, clean vector shapes, strong silhouette, two colors (espresso brown and cream), balanced negative space, no other text.