Card-News Image Generator

Generate content images for card-news using nanobanana MCP. One tool (nanobanana), one output (16:9 editorial PNG files).

---

Purpose

Image Generator produces the visual assets that go into content-image card types. It:

1. Receives image prompts from contents-manager
2. Validates prompts against card-news image requirements
3. Generates images via nanobanana MCP
4. Saves output files for card-news-maker to embed

---

Memory Auto-Load Protocol

1. Load contents-manager output (Visual Asset Plan — provided as input)
   → Image prompts, aspect ratios, save paths
2. Read references/content-image-patterns.md
   → Prompt refinement patterns and style templates
   → If missing: use the built-in prompt formula in Step 1
3. Optional: Load card-news-memory/series-config.md
   → Color theme overrides
   → If missing: use default palette (white/black/yellow #fff3a0)
4. Optional: Load creative-memory/visual-guidelines.md (read-only)
   → If missing: use clean editorial style as default

필수 입력은 contents-manager output뿐이다. 나머지 파일 로드 실패 시 기본값으로 진행.

---

Input

Input	Required	Source
Visual Asset Plan	Yes	contents-manager output
Visual guidelines	Recommended	creative-memory/visual-guidelines.md
Style override	Optional	User preference

---

Process

Step 1: Prompt Review

For each image prompt from the Visual Asset Plan, verify:

Hard requirements checklist:

• Aspect ratio is 16:9
• Prompt explicitly states "no text"
• No watermark/logo instructions
• Subject is clearly described (not vague)
• Style direction is present (editorial, clean, etc.)
• Negative prompt includes standard exclusions

If any requirement is missing, add it before generation.

---

Step 2: Generate Images

Call nanobanana MCP generate_image for each prompt.

Generation parameters:

Parameter	Value
Aspect ratio	16:9
Model	auto (default) or pro for quality-critical images
Max attempts	3 per image

Generation sequence:

1. Call generate_image with the reviewed prompt
2. Evaluate the result:
   • Subject clear and relevant to topic?
   • No unintended text or artifacts?
   • Color palette harmonious with card-news theme?
   • Composition works at 946×576px (the embed area)?
3. If result is unsatisfactory:
   • Adjust prompt (add specificity, strengthen style directive)
   • Retry (up to 3 attempts total)
4. If still unsatisfactory after 3 attempts:
   • Report to user with the best result + prompt used
   • Suggest alternative approaches (different angle, different subject)

Quality checks per image:

• No text visible in the image
• No watermarks or logos
• Clean composition (not overly busy)
• Subject recognizable at card-news display size
• Colors don't clash with the white/black/yellow card-news theme

---

Step 3: Generate Outro Icon (if requested)

If the Visual Asset Plan includes an outro icon prompt:

1. Generate using nanobanana with the specified prompt
2. Target: simple, iconic, suitable for 180×140px display
3. Style: clean illustration or icon, minimal detail

---

Step 4: Save Files

Save each generated image to the project output directory:

File naming convention:

[project]/card-news/[topic]-[YYYY-MM-DD]/img-card-[N].png

Outro icon:

[project]/card-news/[topic]-[YYYY-MM-DD]/outro-icon.png

Example:

card-news/ai-trends-2026-02-25/img-card-2.png
card-news/ai-trends-2026-02-25/img-card-3.png
card-news/ai-trends-2026-02-25/outro-icon.png

---

Step 5: Output Report

# Image Generation Report: [Topic]

> Date: [YYYY-MM-DD]
> Images generated: [N]
> All images: 16:9, no text, editorial style

## Generated Images

### Card [N] — [topic]
- **File**: `[path]`
- **Prompt used**: [final prompt]
- **Model**: [auto/pro/flash]
- **Attempts**: [N]
- **Status**: Success / Partial (notes)

[Repeat for each image]

### Outro Icon
- **File**: `[path]`
- **Status**: [Generated / User-provided / Skipped]

## Production Notes
- [Any quality concerns or user action needed]

---

nanobanana MCP Reference

Parameter	Spec
Tool	generate_image
Models	Gemini 3 Pro (quality) / Gemini 2.5 Flash (speed)
Default model	auto
Max resolution	4K (Pro) / 1024px (Flash)
Prompt limit	8,192 characters

---

What This Skill Does NOT Do

• Plan visual assets → contents-manager creates the prompts
• Write copy → copy-writer
• Render final cards → card-news-maker
• Edit or retouch images → out of scope

Image Generator stays focused: prompts in → PNG files out.