moai-coworker | Image Prompt Builder (text output only)
Overview
Gemini 3 Pro Image (Nano Banana Pro) is Google DeepMind's reasoning-driven image generation and editing model, and Thinking Mode, Perfect Text Rendering, Search Grounding (Google Search integration), Few-Shot Design (up to 14 reference images). The natural language prompt tone works best when it is the tone of a Creative Director directing a scene.
This skill converts a one-line user request into the 5-component structure of the official Google AI for Developers guide:
[Subject + Adjectives] doing [Action] in [Location/Context].
[Composition/Camera]. [Lighting/Atmosphere]. [Style/Media].
[Specific Constraint/Text]
Each component ends with an English sentence and is separated by a period. Keyword-list style works, but result quality is lower.
In particular, this skill:
- 3-model simultaneous output: Gemini 5-component main + GPT-image-2 (6-Block) + Midjourney v8.1 (keywords + parameters)
- Presets + fine-tuning: 4 presets (product shot, portrait, illustration, landscape) × 4 slots
- Thinking vs Fast mode guidance: Thinking for complex compositions and text, Fast for quick exploration (Gemini 3.1 Flash Image)
- Camera hardware specification: Hardware directives that determine visual DNA, such as GoPro, Fujifilm, disposable, iPhone, etc.
Pair skill media-higgsfield-image(Higgsfield MCP — 11 image models including Nano Banana Pro) generates the actual images, and this skill produces only the prompt text .
Trigger keywords
Gemini image prompt nano banana prompt Nano Banana Pro prompt Gemini 3 Pro Image prompt Google AI Studio image Vertex AI image prompt SynthID
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 + 카메라 하드웨어(선택)
↓
[내부] 슬롯 → 5-component 매핑
↓
[내부] 같은 슬롯 → GPT 6-Block + MJ 키워드+파라미터 변환
↓
출력: 3개 모델 프롬프트 코드블록 + 권장 파라미터 + 한국어 해설
Execution rules
Round 1 — Preset selection (required)
AskUserQuestionto have the user select 1 of 4 presets.
The preset slot definitions use a single shared sourceshared by the 3 image prompt builders (gpt-image-2, gemini, midjourney). The source is in the media-gpt-image-2-prompt skill, and each preset file contains tone conversion guides for all three models: GPT, Gemini, and Midjourney.
| Preset | Applicable case | Shared slot source |
|---|---|---|
| Product shot (recommended) | Commerce products, package shots, jewelry and watch close-ups | ${CLAUDE_PLUGIN_ROOT}/skills/media-gpt-image-2-prompt/presets/product-shot.md |
| People and characters | Portraits, personas, advertising models | ${CLAUDE_PLUGIN_ROOT}/skills/media-gpt-image-2-prompt/presets/portrait.md |
| Illustration and art | Card news illustrations, book covers, concept art | ${CLAUDE_PLUGIN_ROOT}/skills/media-gpt-image-2-prompt/presets/illustration.md |
| Landscape and environment | Background images, cinematic backgrounds, travel shots | ${CLAUDE_PLUGIN_ROOT}/skills/media-gpt-image-2-prompt/presets/landscape.md |
Round 2 — Per-Preset Fine-Tuning (3-4 questions)
Follows the slot definitions of the above shared slot source (${CLAUDE_PLUGIN_ROOT}/skills/media-gpt-image-2-prompt/presets/<name>.md). The slot data is used identically by all three models, and this skill applies the Gemini Creative Director tone conversion guide section among them.
Round 3 — Aspect Ratio + Text + Camera Hardware (Optional)
| Aspect Ratio | Gemini Mapping | Purpose |
|---|---|---|
| 1:1 (Recommended) | aspect_ratio="1:1" |
SNS square, general |
| 16:9 | "16:9" |
Wide, YouTube |
| 9:16 | "9:16" |
Reels·Shorts |
| 4:5 | "4:5" |
Instagram Feed |
| 21:9 | "21:9" |
Cinematic Ultra-Wide (Gemini only) |
Gemini additionally supports 3:2, 2:3, 3:4, 4:3, 5:4. Gemini 3.1 Flash Image also additionally supports 1:4, 4:1, 1:8, 8:1.
Camera Hardware Options (Core of the Gemini Creative Director tone):
| Option | Visual DNA |
|---|---|
| Default (DSLR 50mm) | Clean·Neutral·Standard |
| Fujifilm X-T5 | Warm tones, film simulation tone |
| GoPro HERO12 | Wide-angle, action·immersion, slight distortion |
| Disposable film camera | Coarse grain, nostalgic flash, raw mood |
| iPhone 15 Pro | Clean digital, computational photography |
Internal Processing — Slot → 5-Component Mapping
Component 1 — [Subject + Adjectives] doing [Action] in [Location]
Component 2 — [Composition/Camera Angle/Lens/Hardware]
Component 3 — [Lighting/Atmosphere]
Component 4 — [Style/Media]
Component 5 — [Specific Constraint/Text]
Each component is 1-2 English sentences. Separated by periods. For detailed rules, see references/prompt-blocks.md.
Internal Processing — GPT 6-Block + MJ Conversion
Same logic as the paired skills media-gpt-image-2-prompt / media-midjourney-v8-prompt.
Output — 3 Model Code Blocks
## 🎨 생성된 프롬프트 (3개 모델)
### 1) Gemini 3 Pro Image — Nano Banana Pro (메인)
<5-component English sentences>
**권장 파라미터**: `aspect_ratio=1:1`, `resolution=2K`, `mode=Thinking`
**Reference 이미지**: 최대 14개 첨부 가능 (`references/reference-images.md`)
**Search Grounding**: 데이터 시각화·지도·통계 그래프는 활성화 권장
### 2) GPT-image-2 — OpenAI ChatGPT / API
<6-Block natural language paragraph>
**권장 파라미터**: `quality=medium`, `size=1024x1024`, `moderation=auto`
### 3) Midjourney v8.1
<keyword, keyword, ... --ar 1:1 --style raw --hd --q 4 --s 300>
### 📝 한국어 해설
- Gemini는 Creative Director 어조에 가장 잘 반응합니다 (chiaroscuro · golden hour backlighting · three-point softbox 등)
- Thinking Mode는 복잡 구도·텍스트·데이터 시각화에 유리, latency 증가
- 모든 출력 이미지에 SynthID 워터마크 자동 삽입 (imperceptible)
### 🔗 페어 스킬 (실제 이미지 생성)
- `media-higgsfield-image` — Higgsfield MCP 직접 호출 (Nano Banana Pro 포함, 실제 이미지 생성)
- `media-gpt-image-2-prompt` — GPT 어조 프롬프트 빌더 (sibling)
- `media-midjourney-v8-prompt` — MJ 어조 프롬프트 빌더 (sibling)
Text-First Workflow
The 2-step pattern officially recommended by Google when the text to go into the image is long or complex:
- First, refine the text concept through conversation with the model ("recommend a short copy that suits this kind of mood").
- Then put the finalized text verbatim into the image prompt.
This skill automatically recommends using the Text-First pattern in Round 3 when the text length is 30 characters or more.
Usage Examples
Example 1: Product Shot
"Gemini image prompt, product shot of a matte black mug with the text 'MONDAY'"
→ Round 1: Product shot → Round 2: Mug/slate/window/3-4 minutes → Round 3: 1:1 + verbatim "MONDAY" + Fujifilm → Simultaneous output of Gemini main + GPT + MJ.
Example 2: Infographic (Search Grounding)
"Infographic comparing the number of Korean SNS users in 2026 with a nano-banana prompt"
→ Select illustration preset → Round 2 slots → Search Grounding activation guidance + Thinking Mode recommended.
Example 3: Cinematic Landscape 21:9
"Gemini 3 Pro Image cinematic landscape prompt, Han River sunset 21:9"
→ Landscape preset → Round 2 → 21:9 + GoPro wide → Note that only Gemini supports 21:9.
Output Format
| Deliverable | Format | Description |
|---|---|---|
| Gemini 3 Pro Image prompt | English 5-component paragraph | Copy-paste into Google AI Studio / Vertex AI / Gemini app |
| GPT-image-2 prompt | English 6-Block natural language paragraph | Copy-paste into ChatGPT / API |
| Midjourney v8.1 prompt | Keywords + --파라미터 |
Discord /imagine or alpha.midjourney.com |
| Recommended parameters | Per-model aspect/quality/mode | Enter together when configuring API/UI |
| Korean commentary | Markdown | Tone differences · Thinking mode · SynthID · cost cautions |
Cautions
- This skill prompt text only outputs. For actual image generation, use the paired skill.
- All outputs of Gemini 3 Pro Image have an imperceptible SynthID watermark embedded (Google policy, unchangeable).
- Thinking Mode increases latency, but is essential when complex composition, multiple objects, data visualization, or accurate text is required.
- Even when generating fact-based infographics with Search Grounding, results always require separate verification (the model may misinterpret information).
- Input tokens: Gemini 3 Pro Image 65,536 tokens, Flash 131,072 tokens. Output tokens: 32,768 (both).
- When attaching reference images, it is recommended to place key elements (style, character, composition) in the first 2-3, and use the rest for secondary styles.
- Masking edits, day↔night conversion, and multi-image blending sometimes produce unnatural results or artifacts.
Related skills
| Skill | Relationship | Description |
|---|---|---|
| media-gpt-image-2-prompt | sibling | GPT 6-Block tone prompt from the same input |
| media-midjourney-v8-prompt | sibling | MJ keyword+parameter prompt from the same input |
| media-higgsfield-image | after | Actual image generation via direct Higgsfield MCP call (including Nano Banana Pro) |
Sources
Primary recommended sources (official):
- Google AI for Developers — Gemini 3 Pro Image Preview
- Google DeepMind — Gemini 3 Pro Image product page
- Google Cloud Documentation — Gemini 3 Pro Image (Vertex AI)
- Google Cloud Blog — Ultimate Prompting Guide for Nano Banana
- Google AI Studio — Gemini 3 Pro Image (Nano Banana Pro)
Industry references:
- Atlabs AI — Ultimate Nano Banana Pro Prompting Guide 2026
- WaveSpeed Blog — Google Nano Banana Pro: Complete Guide for 2026
- Medium — Testing Gemini 3 Pro Image
Based on the above sources, we derived the 5-component structure, Thinking/Fast mode recommendations, aspect_ratio range, 14 reference images, SynthID policy, 65K/32K token limits, and Search Grounding usage.
References
| File | Load condition |
|---|---|
| references/search-grounding.md | When judging whether to use Search Grounding for images requiring real-time factual data, such as infographics, maps, and statistical graphs |