moai-coworker | Image Prompt Builder (text output only)
Overview
The OpenAI GPT-image-2 model is a reasoning-driven image model that understands natural language in an art-director tone. This skill expands a user's one-line request into the 6-Block structure(Subject → Action → Scene → Composition → Lighting → Style & Text Constraints) recommended by the OpenAI Cookbook, and outputs a prompt that can be copied and pasted directly into ChatGPT or the OpenAI API.
In particular, this skill:
- Simultaneous output for 3 models: Along with the GPT-image-2 main prompt, it also converts the same intent into the tone of Gemini 3 Pro Image (5-component) and Midjourney v8.1 (keywords +
--파라미터) and provides them on a single screen. - Presets + fine-tuning: Lowers the learning curve with 4 presets (product shot, portrait, illustration, landscape), and secures detail with 3-4 fine-tuning questions per preset.
- Guaranteed verbatim text: Uses GPT-image-2's 95%+ text rendering accuracy for text to appear in the image through quotation marks, ALL CAPS, and verbatim instructions.
This skill outputs only prompt text . The user can copy and paste it directly into their desired tool (ChatGPT web, Sora, OpenAI Playground, etc.), or generate images via a direct Higgsfield MCP(Soul) call.
Trigger keywords
GPT image prompt ChatGPT image prompt GPT-image-2 prompt OpenAI image prompt gpt image 2 prompt image prompt builder for GPT OpenAI image creation
Workflow
사용자 자연어 한 줄
↓
[Round 1] AskUserQuestion — 프리셋 선택 (제품샷·인물·일러스트·풍경)
↓
[Round 2] AskUserQuestion — 프리셋별 미세조정 (3~4 슬롯)
↓
[Round 3] AskUserQuestion — 화면비 + 이미지 내 텍스트 유무
↓
[내부] 슬롯 → 6-Block 매핑 (Subject·Action·Scene·Composition·Lighting·Style&Text)
↓
[내부] 같은 슬롯 → Gemini 5-component 변환 + MJ 키워드+파라미터 변환
↓
출력: 3개 모델 프롬프트 코드블록 + 권장 파라미터 + 한국어 해설
Execution rules
Round 1 — Preset selection (required)
AskUserQuestionand receive a selection of 1 out of 4 presets.
| Preset | Applicable cases | references |
|---|---|---|
| Product shot (recommended) | Commerce product photos, package shots, jewelry and watch close-ups | presets/product-shot.md |
| People and characters | Portraits, persona illustrations, advertising models | presets/portrait.md |
| Illustration and art | Card news illustrations, book covers, concept art | presets/illustration.md |
| Landscape and environment | Background images, spatial photos, cinematic backgrounds | presets/landscape.md |
The selection result determines the question set for Round 2.
Round 2 — Preset-specific fine-tuning (3-4 questions)
The question set defined in the selected preset's presets/<name>.mdis iterated through with AskUserQuestion. Each question has 4 options + Other, and the first option is labeled (권장) .
Product shot example:
- Product and material (e.g., matte black ceramic mug, wood tray + leather laptop sleeve)
- Background·Surface (e.g., wet slate counter, beige linen, white studio)
- Lighting·Time of day (e.g., window side light sunrise, studio softbox, golden hour)
- Camera angle·Lens (e.g., 3/4 angle 50mm, top-down 35mm, macro 100mm)
Slots of other presets are presets/portrait.md, presets/illustration.md, presets/landscape.mddefined in .
Round 3 — Aspect ratio + Text (1-2 questions)
| Aspect ratio options | Purpose | Mapping by model |
|---|---|---|
| 1:1 (recommended) | SNS square, general | GPT 1024x1024, Gemini 1:1, MJ --ar 1:1 |
| 16:9 | YouTube thumbnail, wide | GPT 1536x864, Gemini 16:9, MJ --ar 16:9 |
| 9:16 | Instagram Reels·Shorts | GPT 768x1344, Gemini 9:16, MJ --ar 9:16 |
| 4:5 | Instagram feed | GPT 1024x1280, Gemini 4:5, MJ --ar 4:5 |
If there is text in the image, the exact string is requested through a separate question (to preserve verbatim).
Internal processing — Slot → 6-Block mapping
The collected slots are written out as one natural-language paragraph in the tone of the OpenAI Cookbook's 6-Block. Not a keyword list format, but a tone as if an art director is explainingis required. For detailed rules, refer to references/prompt-blocks.md.
| Block | Slot mapping | Tone example |
|---|---|---|
| Subject | The preset's core subject slot | "A matte black ceramic coffee mug with subtle ridge texture" |
| Action | Action·Placement (static if absent) | "sitting on a wet slate countertop next to a folded linen napkin" |
| Scene | Background·Context slot | "in a minimalist Scandinavian kitchen at sunrise" |
| Composition | Angle·Lens slot | "three-quarter angle, 50mm lens, shallow depth of field" |
| Lighting | Lighting slot | "soft directional window light, cool morning tone" |
| Style & Text | Style + Round 3 text | "editorial product photography, natural film grain. The mug has "MONDAY" printed in thin uppercase sans-serif on the side, perfectly legible." |
Text verbatim rules are references/text-rendering.mdorganized in .
Internal processing — Gemini 5-component conversion
The same slots are converted into Gemini 3 Pro Image's 5-component English sentence format (each component separated by a period, Creative Director tone).
[Subject + Adjectives] doing [Action] in [Location/Context].
[Composition/Camera]. [Lighting/Atmosphere]. [Style/Media].
[Specific Constraint/Text]
Internal processing — Midjourney v8 keyword+parameter conversion
The same slots are converted into a comma-separated keyword + --파라미터 format.
[subject], [scene keywords], [composition], [lighting], [style] --ar W:H --style raw --hd --q 4 --s 0~1000 [--sref CODE|URL --sw N] [--oref URL --cw 0~100] [--p PROFILE] [--no NEGATIVE]
The parameter usage guide is references/parameter-cheatsheet.mdorganized in . In particular, the user is notified of the --cw 100(default) pitfall and the 4x cost caution of --hd --q 4 --sref.
Output — 3 model code blocks + Korean commentary + paired skill guide
## 🎨 생성된 프롬프트 (3개 모델)
### 1) GPT-image-2 — OpenAI ChatGPT / API
<6-Block natural language paragraph>
**권장 파라미터**: `quality=medium`, `size=1024x1024`, `moderation=auto`
**편집 시**: `references/editing-patterns.md`의 Change/Preserve/Constraints 2열 로직 적용
### 2) Gemini 3 Pro Image — Nano Banana Pro
<5-component English sentence>
**권장 파라미터**: `aspect_ratio=1:1`, `resolution=2K`, `mode=Thinking`
### 3) Midjourney v8.1
<keyword, keyword, ... --ar 1:1 --style raw --hd --q 4 --s 300>
**비용 주의**: `--hd` + `--q 4` 조합은 4x GPU 시간
### 📝 한국어 해설
- 3개는 같은 장면이지만 어조가 다릅니다 (자연어 단락 / 5-component 문장 / 키워드+파라미터)
- 텍스트 정확도: GPT 95%+, Gemini 우수, MJ V8에서 개선됨
### 🔗 실제 이미지 생성
- **Higgsfield MCP**(Soul) — 시네마틱 이미지·캐릭터 단일 통합 직접 호출
- ChatGPT 웹·Sora·OpenAI Playground 직접 복붙
Editing workflow (optional)
If the user requests "edit this image," enter edit mode instead of Round 2 and apply GPT-image-2's two-column logic.
Change: <바꿀 요소를 한 문장으로>
Preserve: <face·pose·lighting·framing·background·geometry·text·layout 중 잠금할 항목 나열>
Constraints: <no extra objects, no redesign, no logo drift, no watermark>
For details, references/editing-patterns.md.
Usage examples
Example 1: One-line request for a product shot
"Make me an image prompt for GPT. A product shot of a matte black mug with the letters 'MONDAY'"
→ Round 1: Select product shot → Round 2: Enter mug/slate counter/window lighting/3-4 angle slots → Round 3: 1:1 + "MONDAY" verbatim → Output prompts for 3 models.
Example 2: Person prompt for Instagram Reels
"ChatGPT image prompt, lifestyle shot of a woman in her 30s looking at a laptop in a cafe, 9:16"
→ Round 1: Character selection → Round 2: Woman in her 30s/cafe·props/expression·action/background → Round 3: 9:16 + no text → 3 model prompts + Sora video conversion hints.
Example 3: Edit mode
"GPT prompt to change only the mug color to red in an existing mug image"
→ Enter edit mode → Collect Change/Preserve/Constraints slots → Output only the prompt for GPT-image-2 editing (Gemini and MJ use different editing mechanisms, so separate guidance is provided).
Output format
| Deliverable | Format | Description |
|---|---|---|
| GPT-image-2 prompt | English natural-language paragraph (1) | ChatGPT / API prompt field copy-paste |
| Gemini 3 Pro Image prompt | English 5-component paragraph | Google AI Studio / Vertex AI prompt field copy-paste |
| Midjourney v8.1 prompt | Keywords + --파라미터 |
Discord /imagine or alpha.midjourney.com input |
| Recommended parameter table | quality/size/aspect per model | Set together when calling the API |
| Korean commentary | Markdown | Tone differences·text accuracy·cost cautions |
Cautions
- This skill prompt text only outputs. For actual image generation, use the paired skill.
- GPT-image-2 4K output is classified by OpenAI as experimental. For general production, 1024 or 2K is recommended.
- Midjourney
--cw 100(Default) Pitfall: the lighting and style of the reference image may also be inherited and conflict with the new prompt. To take only the face,--cw 0~30. - In images with text, GPT-image-2 can also handle non-Latin scripts such as Korean, Japanese, Chinese, Hindi, and Bengali, but for accuracy, quotation marks + verbatim instruction + font weight·color·position must be specified.
- All prompts are used under the user's responsibility, and this skill bears no responsibility for copyright, portrait rights, or brand usage.
- A SynthID watermark is imperceptibly embedded in Gemini 3 Pro Image output (Google policy).
Related skills
| Skill | Relationship | Description |
|---|---|---|
| media-gemini-3-image-prompt | sibling | Produces a Gemini tone-optimized prompt from the same input |
| media-midjourney-v8-prompt | sibling | Produces an MJ keyword+parameter prompt from the same input |
| media-codex-image | after | Generates images with codex CLI (gpt-image-2, no OAuth·API key required) using the generated prompt |
| media-higgsfield-image | after | Directly generates images with Higgsfield MCP using the generated prompt (including GPT Image 2·Nano Banana Pro) |
Sources
Primary recommended sources (official):
- OpenAI Cookbook — GPT Image Generation Models Prompting Guide
- openai/openai-cookbook GitHub — image-gen-models-prompting-guide.ipynb
Industry references:
Based on the above sources, the 6-Block structure, recommended quality/size values, edit two-column logic, and text verbatim rules were derived.