Xiaohongshu Image Workflow

⚠️ Core Principle: Propose First, Then Generate

Never skip the design proposal and jump straight to generating. The correct workflow:

Understand content → Design proposal (2–3 directions) → User selection → Generate → Preview confirmation → Upload

Huashu's Design Aesthetic Profile

What He Likes

• Texture and warmth — Organic elements such as paper folds, handwritten strokes, stamps, washi tape
• Handwritten / calligraphic fonts — Notebook feel; especially effective for personal sharing scenarios
• Text filling the canvas — Xiaohongshu is a mobile vertical screen; text must be large enough to read at a glance
• Core element emphasis — Key numbers / keywords as visual heroes (3× enlarged, colour-changed, decorated)
• Clear structure — Main title > subtitle > list; distinct hierarchy
• Warm colour tones — Cream, warm orange, warm gold perform well

What He Dislikes

• HTML screenshots — Too flat, feels like a PowerPoint template, soulless (only use as precise data table fallback)
• Cyber-neon / dark blue backgrounds — #0D1117 and similar dark blues are an aesthetic no-go zone
• Signature / watermark — Cover images must not include "Huashu" or "@Huashu"
• Excessive blank space — Better to be fuller than sparse and empty

Proven Good Styles

| Style | Performance | Best For |

param($m) $inner = $m.Groups[1].Value

Split by | and fix each cell separator

$cells = $inner -split '\|' $fixedCells = $cells | ForEach-Object { $cell = --- name: huashu-xhs-image description: | Generate high-quality images for Xiaohongshu (Little Red Book) posts. Default path is AI-generated (Gemini); HTML is only used as fallback for precise data tables. Use when the user mentions "Xiaohongshu images", "Xiaohongshu cover", "Xiaohongshu pictures", "make a Xiaohongshu image", or "post illustrations".

Xiaohongshu Image Workflow

⚠️ Core Principle: Propose First, Then Generate

Never skip the design proposal and jump straight to generating. The correct workflow:

Understand content → Design proposal (2–3 directions) → User selection → Generate → Preview confirmation → Upload

Huashu's Design Aesthetic Profile

What He Likes

• Texture and warmth — Organic elements such as paper folds, handwritten strokes, stamps, washi tape
• Handwritten / calligraphic fonts — Notebook feel; especially effective for personal sharing scenarios
• Text filling the canvas — Xiaohongshu is a mobile vertical screen; text must be large enough to read at a glance
• Core element emphasis — Key numbers / keywords as visual heroes (3× enlarged, colour-changed, decorated)
• Clear structure — Main title > subtitle > list; distinct hierarchy
• Warm colour tones — Cream, warm orange, warm gold perform well

What He Dislikes

• HTML screenshots — Too flat, feels like a PowerPoint template, soulless (only use as precise data table fallback)
• Cyber-neon / dark blue backgrounds — #0D1117 and similar dark blues are an aesthetic no-go zone
• Signature / watermark — Cover images must not include "Huashu" or "@Huashu"
• Excessive blank space — Better to be fuller than sparse and empty

Proven Good Styles

Core Parameters

Step 1: Understand the Content

Read the user-provided content and quickly extract:

• Theme: What is this post about?
• Core keywords: Which words / numbers need to be visual heroes?
• Mood / tone: Suspense? Informative? Warm? Impactful?
• Image count and type: Single cover / carousel set / infographic?

No need to show the analysis to the user — proceed directly to Step 2.

Step 2: Design Proposal ✅ Must Wait for User Selection

This is the most critical step in the entire workflow. Never skip it.

Proposal Format

Present 2–3 design directions to the user, each including:

### Direction A: [Style Name]
- Visual style: [One sentence describing the visual feel, e.g., "warm cream notebook paper + brush calligraphy headline + hand-drawn icons"]
- Colour palette: [Background + primary colour + accent colour]
- Copy layout: [Which text is enlarged as the hero, which is subtitle, overall arrangement]
- Mood: [The first feeling the viewer gets]

Proposal Principles

1. Each direction must be clearly different (style, mood, and colour must differ on at least one dimension)
2. Mark recommendations (explain why a particular direction is recommended based on the content)
3. Be specific about copy (not "title enlarged" — say "the number '28' as a 200px hero element in orange")
4. No more than 3 directions (too many choices makes selecting harder)

Proposal Example

Direction A: Hand-drawn Notebook Style (Recommended)

>

- Visual style: Cream grid paper background + brush calligraphy headline + hand-drawn tech icons - Colour palette: Background #FDF6EC + Primary #D97706 (warm orange) + emphasis circles - Copy layout: "Ali C4 Building" and "A Bunch of People from Guangdong Arrived" fill the upper half as heroes; "AI Open-Source Champion" in orange as the visual anchor; bottom right "Qianwen App" stamp - Mood: Friendly, authentic, like a friend sharing insider news

Direction B: Dark Gold Reveal Style

>

- Visual style: Dark frosted background + gold large text + badge decoration - Colour palette: Background #1A1A1A + Primary #E2B714 (gold) + white as support - Copy layout: "Global AI Open-Source Champion" in gold fills the canvas; "Big Tech Insider" gold badge at top; white subtitle below - Mood: Impactful, insider feel, authoritative

Only proceed to Step 3 after the user has made a selection. The user might:

• Select one directly → Proceed to generation
• Ask for a mix / adjustments → Modify the direction and re-confirm
• Be unhappy with all options → Ask for preferences and re-propose

Step 3: Generate Images

Build the Prompt

Based on the user's chosen direction, build a complete prompt.

Prompt template:

Create a [style] cover for a Xiaohongshu post. 3:4 portrait aspect ratio, 1080x1440 pixels, high quality rendering.

VISUAL STYLE: [Expand from the visual style description in the proposal]
COLOR PALETTE: [Specific colour description]
TYPOGRAPHY: text fills most of the canvas, oversized bold typography, clear visual hierarchy.

TEXT TO RENDER:
- [Main title — hero element, visually dominant]
- [Subtitle]
- [Other text elements]

The word/number "[core keyword]" is visually dominant, 3x larger than other text, with decorative emphasis.

IMPORTANT: Do NOT include any personal signature, watermark, or author name.

[1–2 sentences describing the mood / atmosphere of the image]

Key prompt keywords (add as needed):

• Large text → text fills most of the canvas, oversized bold typography
• Core emphasis → the word/number "XX" is visually dominant, 3x larger than other text, with decorative emphasis
• Handwriting → handwritten style Chinese text / brush calligraphy lettering
• Paper texture → warm cream paper texture with subtle grid lines, notebook page feel
• Clear hierarchy → clear visual hierarchy with distinct heading, subheading, and list levels
• No signature → Do NOT include any personal signature, watermark, or author name

Two Generation Paths (Always Provide Both)

Always provide both paths so the user can compare. HTML is zero cost, so multiple colour / layout variants can be generated per direction. AI path — 1 image per direction is sufficient.

File Organisation (Must Follow)

When generating multiple versions, all related files (PNG + HTML source) go in the same subfolder:

[Article directory]/
├── article.md
└── [article-short-name]-xhs-images/          ← Subfolder
    ├── A-notebook-AI.png
    ├── A1-notebook-HTML-warm.png
    ├── A1-notebook-HTML-warm.html
    ├── B-newspaper-AI.png
    └── ...

Naming convention: [Direction letter][Variant number]-[Style name]-[Path AI/HTML]-[Variant description].png

• Direction letter: A/B/C (matching the design proposals)
• AI path — no variant number: A-notebook-AI.png
• HTML variants — with variant number: A1-notebook-HTML-warm.png, A2-notebook-HTML-green.png
• Folder name uses the article keyword: [keyword]-xhs-images/

AI Generation Command

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run /path/to/skills/xhs-image/scripts/generate_image.py \
  --prompt "[full prompt]" \
  --filename "[direction]-[style]-AI.png" \
  --resolution 2K

After generation, move to the images subfolder. Sets can be generated in parallel (run_in_background=true).

Step 4: Preview Confirmation

Browser Preview (Required)

After generation, open all images with the open command so the user can compare side by side:

open "[image path 1]" "[image path 2]" "[image path 3]"

Inline Preview

Also display generation results inline using the Read tool.

Basic checks:

1. Are the Chinese characters rendered correctly?
2. Is the ratio 3:4 portrait orientation?
3. Does the style match the selected direction?
4. Does it include any signature or watermark?

Design Review (Required)

Score each image on two dimensions (out of 10) and provide optimisation directions:

Review output format:

• Per image: Overall score + 1-sentence core evaluation + 1 optimisation direction
• Final summary ranking table with recommendations highlighted
• The user can decide whether to adopt the optimisation suggestions

User feedback handling:

• Satisfied → Step 5 upload
• Text errors → Use HTML rendering as fallback for that image
• Style wrong → Adjust prompt and regenerate
• Major direction change → Return to Step 2 for a new proposal

Step 5: Upload to Image Host

python3 /path/to/tools/upload_image.py "[image path]"

Returns a permanent ImgBB link.

HTML Screenshot Path (For AI text rendering failures, precise data tables, or user-requested comparisons)

npx playwright screenshot "file:///path/to/card.html" output.png \
  --viewport-size=1080,1440 --wait-for-timeout=1000

HTML template requirements:

• Canvas: width: 1080px; height: 1440px
• Font: font-family: "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei", sans-serif
• Safe zone: 80px top/bottom, 60px left/right

Quick Reference

Chinese Text Rendering Limits (AI Path)

• Main title: ≤ 7 characters
• Subtitle: ≤ 15 characters
• Body text: ≤ 20 characters per line
• Verify every image individually

Huashu Tech Account Colour Palette

Golden Rules

• Title: large, bold, prominent (occupies 30–50% of canvas)
• Core numbers / keywords as visual emphasis (enlarged, coloured, decorated)
• Cover information density → sparks curiosity
• Carousel sets maintain consistent style
• Vertical 3:4, making full use of screen real estate
• No signature / watermark

Related Skills

Reference Files

• references/style-gallery.md — Full style library and prompt templates
• references/design-guidelines.md — Xiaohongshu platform design specifications

By Huashu | AI Native Coder · Independent Developer WeChat Official Account "Huashu" | 300k+ followers | AI Tools & Productivity Flagship products: Kitty Catchlight (AppStore Paid Rankings Top 1) · Play DeepSeek with One Book

.Trim() if ($cell -match '^:?-+:? | Hand-drawn notebook (warm paper + calligraphy + hand-drawn icons) | ⭐⭐⭐⭐⭐ | Tutorials, tips, personal sharing | | Dark gold poster (dark background + gold large text) | ⭐⭐⭐⭐ | Product launches, impactful titles (needs strong content) | | Minimalist infographic (light background + large numbers + clean hierarchy) | ⭐⭐⭐⭐ | Data-driven, comparisons |

Core Parameters

Step 1: Understand the Content

Read the user-provided content and quickly extract:

• Theme: What is this post about?
• Core keywords: Which words / numbers need to be visual heroes?
• Mood / tone: Suspense? Informative? Warm? Impactful?
• Image count and type: Single cover / carousel set / infographic?

No need to show the analysis to the user — proceed directly to Step 2.

Step 2: Design Proposal ✅ Must Wait for User Selection

This is the most critical step in the entire workflow. Never skip it.

Proposal Format

Present 2–3 design directions to the user, each including:

### Direction A: [Style Name]
- Visual style: [One sentence describing the visual feel, e.g., "warm cream notebook paper + brush calligraphy headline + hand-drawn icons"]
- Colour palette: [Background + primary colour + accent colour]
- Copy layout: [Which text is enlarged as the hero, which is subtitle, overall arrangement]
- Mood: [The first feeling the viewer gets]

Proposal Principles

1. Each direction must be clearly different (style, mood, and colour must differ on at least one dimension)
2. Mark recommendations (explain why a particular direction is recommended based on the content)
3. Be specific about copy (not "title enlarged" — say "the number '28' as a 200px hero element in orange")
4. No more than 3 directions (too many choices makes selecting harder)

Proposal Example

Direction A: Hand-drawn Notebook Style (Recommended)

>

- Visual style: Cream grid paper background + brush calligraphy headline + hand-drawn tech icons - Colour palette: Background #FDF6EC + Primary #D97706 (warm orange) + emphasis circles - Copy layout: "Ali C4 Building" and "A Bunch of People from Guangdong Arrived" fill the upper half as heroes; "AI Open-Source Champion" in orange as the visual anchor; bottom right "Qianwen App" stamp - Mood: Friendly, authentic, like a friend sharing insider news

Direction B: Dark Gold Reveal Style

>

- Visual style: Dark frosted background + gold large text + badge decoration - Colour palette: Background #1A1A1A + Primary #E2B714 (gold) + white as support - Copy layout: "Global AI Open-Source Champion" in gold fills the canvas; "Big Tech Insider" gold badge at top; white subtitle below - Mood: Impactful, insider feel, authoritative

Only proceed to Step 3 after the user has made a selection. The user might:

• Select one directly → Proceed to generation
• Ask for a mix / adjustments → Modify the direction and re-confirm
• Be unhappy with all options → Ask for preferences and re-propose

Step 3: Generate Images

Build the Prompt

Based on the user's chosen direction, build a complete prompt.

Prompt template:

Create a [style] cover for a Xiaohongshu post. 3:4 portrait aspect ratio, 1080x1440 pixels, high quality rendering.

VISUAL STYLE: [Expand from the visual style description in the proposal]
COLOR PALETTE: [Specific colour description]
TYPOGRAPHY: text fills most of the canvas, oversized bold typography, clear visual hierarchy.

TEXT TO RENDER:
- [Main title — hero element, visually dominant]
- [Subtitle]
- [Other text elements]

The word/number "[core keyword]" is visually dominant, 3x larger than other text, with decorative emphasis.

IMPORTANT: Do NOT include any personal signature, watermark, or author name.

[1–2 sentences describing the mood / atmosphere of the image]

Key prompt keywords (add as needed):

• Large text → text fills most of the canvas, oversized bold typography
• Core emphasis → the word/number "XX" is visually dominant, 3x larger than other text, with decorative emphasis
• Handwriting → handwritten style Chinese text / brush calligraphy lettering
• Paper texture → warm cream paper texture with subtle grid lines, notebook page feel
• Clear hierarchy → clear visual hierarchy with distinct heading, subheading, and list levels
• No signature → Do NOT include any personal signature, watermark, or author name

Two Generation Paths (Always Provide Both)

Always provide both paths so the user can compare. HTML is zero cost, so multiple colour / layout variants can be generated per direction. AI path — 1 image per direction is sufficient.

File Organisation (Must Follow)

When generating multiple versions, all related files (PNG + HTML source) go in the same subfolder:

[Article directory]/
├── article.md
└── [article-short-name]-xhs-images/          ← Subfolder
    ├── A-notebook-AI.png
    ├── A1-notebook-HTML-warm.png
    ├── A1-notebook-HTML-warm.html
    ├── B-newspaper-AI.png
    └── ...

Naming convention: [Direction letter][Variant number]-[Style name]-[Path AI/HTML]-[Variant description].png

• Direction letter: A/B/C (matching the design proposals)
• AI path — no variant number: A-notebook-AI.png
• HTML variants — with variant number: A1-notebook-HTML-warm.png, A2-notebook-HTML-green.png
• Folder name uses the article keyword: [keyword]-xhs-images/

AI Generation Command

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run /path/to/skills/xhs-image/scripts/generate_image.py \
  --prompt "[full prompt]" \
  --filename "[direction]-[style]-AI.png" \
  --resolution 2K

After generation, move to the images subfolder. Sets can be generated in parallel (run_in_background=true).

Step 4: Preview Confirmation

Browser Preview (Required)

After generation, open all images with the open command so the user can compare side by side:

open "[image path 1]" "[image path 2]" "[image path 3]"

Inline Preview

Also display generation results inline using the Read tool.

Basic checks:

1. Are the Chinese characters rendered correctly?
2. Is the ratio 3:4 portrait orientation?
3. Does the style match the selected direction?
4. Does it include any signature or watermark?

Design Review (Required)

Score each image on two dimensions (out of 10) and provide optimisation directions:

Review output format:

• Per image: Overall score + 1-sentence core evaluation + 1 optimisation direction
• Final summary ranking table with recommendations highlighted
• The user can decide whether to adopt the optimisation suggestions

User feedback handling:

• Satisfied → Step 5 upload
• Text errors → Use HTML rendering as fallback for that image
• Style wrong → Adjust prompt and regenerate
• Major direction change → Return to Step 2 for a new proposal

Step 5: Upload to Image Host

python3 /path/to/tools/upload_image.py "[image path]"

Returns a permanent ImgBB link.

HTML Screenshot Path (For AI text rendering failures, precise data tables, or user-requested comparisons)

npx playwright screenshot "file:///path/to/card.html" output.png \
  --viewport-size=1080,1440 --wait-for-timeout=1000

HTML template requirements:

• Canvas: width: 1080px; height: 1440px
• Font: font-family: "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei", sans-serif
• Safe zone: 80px top/bottom, 60px left/right

Quick Reference

Chinese Text Rendering Limits (AI Path)

• Main title: ≤ 7 characters
• Subtitle: ≤ 15 characters
• Body text: ≤ 20 characters per line
• Verify every image individually

Huashu Tech Account Colour Palette

Golden Rules

• Title: large, bold, prominent (occupies 30–50% of canvas)
• Core numbers / keywords as visual emphasis (enlarged, coloured, decorated)
• Cover information density → sparks curiosity
• Carousel sets maintain consistent style
• Vertical 3:4, making full use of screen real estate
• No signature / watermark

Related Skills

Reference Files

• references/style-gallery.md — Full style library and prompt templates
• references/design-guidelines.md — Xiaohongshu platform design specifications

By Huashu | AI Native Coder · Independent Developer WeChat Official Account "Huashu" | 300k+ followers | AI Tools & Productivity Flagship products: Kitty Catchlight (AppStore Paid Rankings Top 1) · Play DeepSeek with One Book

) { " $cell " } else { " $cell " } } '|' + ($fixedCells -join '|') + '|'

| Hand-drawn notebook (warm paper + calligraphy + hand-drawn icons) | ⭐⭐⭐⭐⭐ | Tutorials, tips, personal sharing | | Dark gold poster (dark background + gold large text) | ⭐⭐⭐⭐ | Product launches, impactful titles (needs strong content) | | Minimalist infographic (light background + large numbers + clean hierarchy) | ⭐⭐⭐⭐ | Data-driven, comparisons |

Core Parameters

| Parameter | Value |

param($m) $inner = $m.Groups[1].Value

Split by | and fix each cell separator

$cells = $inner -split '\|' $fixedCells = $cells | ForEach-Object { $cell = --- name: huashu-xhs-image description: | Generate high-quality images for Xiaohongshu (Little Red Book) posts. Default path is AI-generated (Gemini); HTML is only used as fallback for precise data tables. Use when the user mentions "Xiaohongshu images", "Xiaohongshu cover", "Xiaohongshu pictures", "make a Xiaohongshu image", or "post illustrations".

Xiaohongshu Image Workflow

⚠️ Core Principle: Propose First, Then Generate

Never skip the design proposal and jump straight to generating. The correct workflow:

Understand content → Design proposal (2–3 directions) → User selection → Generate → Preview confirmation → Upload

Huashu's Design Aesthetic Profile

What He Likes

• Texture and warmth — Organic elements such as paper folds, handwritten strokes, stamps, washi tape
• Handwritten / calligraphic fonts — Notebook feel; especially effective for personal sharing scenarios
• Text filling the canvas — Xiaohongshu is a mobile vertical screen; text must be large enough to read at a glance
• Core element emphasis — Key numbers / keywords as visual heroes (3× enlarged, colour-changed, decorated)
• Clear structure — Main title > subtitle > list; distinct hierarchy
• Warm colour tones — Cream, warm orange, warm gold perform well

What He Dislikes

• HTML screenshots — Too flat, feels like a PowerPoint template, soulless (only use as precise data table fallback)
• Cyber-neon / dark blue backgrounds — #0D1117 and similar dark blues are an aesthetic no-go zone
• Signature / watermark — Cover images must not include "Huashu" or "@Huashu"
• Excessive blank space — Better to be fuller than sparse and empty

Proven Good Styles

Core Parameters

Step 1: Understand the Content

Read the user-provided content and quickly extract:

• Theme: What is this post about?
• Core keywords: Which words / numbers need to be visual heroes?
• Mood / tone: Suspense? Informative? Warm? Impactful?
• Image count and type: Single cover / carousel set / infographic?

No need to show the analysis to the user — proceed directly to Step 2.

Step 2: Design Proposal ✅ Must Wait for User Selection

This is the most critical step in the entire workflow. Never skip it.

Proposal Format

Present 2–3 design directions to the user, each including:

### Direction A: [Style Name]
- Visual style: [One sentence describing the visual feel, e.g., "warm cream notebook paper + brush calligraphy headline + hand-drawn icons"]
- Colour palette: [Background + primary colour + accent colour]
- Copy layout: [Which text is enlarged as the hero, which is subtitle, overall arrangement]
- Mood: [The first feeling the viewer gets]

Proposal Principles

1. Each direction must be clearly different (style, mood, and colour must differ on at least one dimension)
2. Mark recommendations (explain why a particular direction is recommended based on the content)
3. Be specific about copy (not "title enlarged" — say "the number '28' as a 200px hero element in orange")
4. No more than 3 directions (too many choices makes selecting harder)

Proposal Example

Direction A: Hand-drawn Notebook Style (Recommended)

>

- Visual style: Cream grid paper background + brush calligraphy headline + hand-drawn tech icons - Colour palette: Background #FDF6EC + Primary #D97706 (warm orange) + emphasis circles - Copy layout: "Ali C4 Building" and "A Bunch of People from Guangdong Arrived" fill the upper half as heroes; "AI Open-Source Champion" in orange as the visual anchor; bottom right "Qianwen App" stamp - Mood: Friendly, authentic, like a friend sharing insider news

Direction B: Dark Gold Reveal Style

>

- Visual style: Dark frosted background + gold large text + badge decoration - Colour palette: Background #1A1A1A + Primary #E2B714 (gold) + white as support - Copy layout: "Global AI Open-Source Champion" in gold fills the canvas; "Big Tech Insider" gold badge at top; white subtitle below - Mood: Impactful, insider feel, authoritative

Only proceed to Step 3 after the user has made a selection. The user might:

• Select one directly → Proceed to generation
• Ask for a mix / adjustments → Modify the direction and re-confirm
• Be unhappy with all options → Ask for preferences and re-propose

Step 3: Generate Images

Build the Prompt

Based on the user's chosen direction, build a complete prompt.

Prompt template:

Create a [style] cover for a Xiaohongshu post. 3:4 portrait aspect ratio, 1080x1440 pixels, high quality rendering.

VISUAL STYLE: [Expand from the visual style description in the proposal]
COLOR PALETTE: [Specific colour description]
TYPOGRAPHY: text fills most of the canvas, oversized bold typography, clear visual hierarchy.

TEXT TO RENDER:
- [Main title — hero element, visually dominant]
- [Subtitle]
- [Other text elements]

The word/number "[core keyword]" is visually dominant, 3x larger than other text, with decorative emphasis.

IMPORTANT: Do NOT include any personal signature, watermark, or author name.

[1–2 sentences describing the mood / atmosphere of the image]

Key prompt keywords (add as needed):

• Large text → text fills most of the canvas, oversized bold typography
• Core emphasis → the word/number "XX" is visually dominant, 3x larger than other text, with decorative emphasis
• Handwriting → handwritten style Chinese text / brush calligraphy lettering
• Paper texture → warm cream paper texture with subtle grid lines, notebook page feel
• Clear hierarchy → clear visual hierarchy with distinct heading, subheading, and list levels
• No signature → Do NOT include any personal signature, watermark, or author name

Two Generation Paths (Always Provide Both)

Always provide both paths so the user can compare. HTML is zero cost, so multiple colour / layout variants can be generated per direction. AI path — 1 image per direction is sufficient.

File Organisation (Must Follow)

When generating multiple versions, all related files (PNG + HTML source) go in the same subfolder:

[Article directory]/
├── article.md
└── [article-short-name]-xhs-images/          ← Subfolder
    ├── A-notebook-AI.png
    ├── A1-notebook-HTML-warm.png
    ├── A1-notebook-HTML-warm.html
    ├── B-newspaper-AI.png
    └── ...

Naming convention: [Direction letter][Variant number]-[Style name]-[Path AI/HTML]-[Variant description].png

• Direction letter: A/B/C (matching the design proposals)
• AI path — no variant number: A-notebook-AI.png
• HTML variants — with variant number: A1-notebook-HTML-warm.png, A2-notebook-HTML-green.png
• Folder name uses the article keyword: [keyword]-xhs-images/

AI Generation Command

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run /path/to/skills/xhs-image/scripts/generate_image.py \
  --prompt "[full prompt]" \
  --filename "[direction]-[style]-AI.png" \
  --resolution 2K

After generation, move to the images subfolder. Sets can be generated in parallel (run_in_background=true).

Step 4: Preview Confirmation

Browser Preview (Required)

After generation, open all images with the open command so the user can compare side by side:

open "[image path 1]" "[image path 2]" "[image path 3]"

Inline Preview

Also display generation results inline using the Read tool.

Basic checks:

1. Are the Chinese characters rendered correctly?
2. Is the ratio 3:4 portrait orientation?
3. Does the style match the selected direction?
4. Does it include any signature or watermark?

Design Review (Required)

Score each image on two dimensions (out of 10) and provide optimisation directions:

Review output format:

• Per image: Overall score + 1-sentence core evaluation + 1 optimisation direction
• Final summary ranking table with recommendations highlighted
• The user can decide whether to adopt the optimisation suggestions

User feedback handling:

• Satisfied → Step 5 upload
• Text errors → Use HTML rendering as fallback for that image
• Style wrong → Adjust prompt and regenerate
• Major direction change → Return to Step 2 for a new proposal

Step 5: Upload to Image Host

python3 /path/to/tools/upload_image.py "[image path]"

Returns a permanent ImgBB link.

HTML Screenshot Path (For AI text rendering failures, precise data tables, or user-requested comparisons)

npx playwright screenshot "file:///path/to/card.html" output.png \
  --viewport-size=1080,1440 --wait-for-timeout=1000

HTML template requirements:

• Canvas: width: 1080px; height: 1440px
• Font: font-family: "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei", sans-serif
• Safe zone: 80px top/bottom, 60px left/right

Quick Reference

Chinese Text Rendering Limits (AI Path)

• Main title: ≤ 7 characters
• Subtitle: ≤ 15 characters
• Body text: ≤ 20 characters per line
• Verify every image individually

Huashu Tech Account Colour Palette

Golden Rules

• Title: large, bold, prominent (occupies 30–50% of canvas)
• Core numbers / keywords as visual emphasis (enlarged, coloured, decorated)
• Cover information density → sparks curiosity
• Carousel sets maintain consistent style
• Vertical 3:4, making full use of screen real estate
• No signature / watermark

Related Skills

Reference Files

• references/style-gallery.md — Full style library and prompt templates
• references/design-guidelines.md — Xiaohongshu platform design specifications

By Huashu | AI Native Coder · Independent Developer WeChat Official Account "Huashu" | 300k+ followers | AI Tools & Productivity Flagship products: Kitty Catchlight (AppStore Paid Rankings Top 1) · Play DeepSeek with One Book

.Trim() if ($cell -match '^:?-+:? | Standard dimensions | 1080 × 1440 px (3:4) | | AI generation resolution | --resolution 2K | | AI Prompt aspect ratio declaration | 3:4 portrait aspect ratio, 1080x1440 pixels | | HTML viewport (fallback) | --viewport-size=1080,1440 |

Step 1: Understand the Content

Read the user-provided content and quickly extract:

• Theme: What is this post about?
• Core keywords: Which words / numbers need to be visual heroes?
• Mood / tone: Suspense? Informative? Warm? Impactful?
• Image count and type: Single cover / carousel set / infographic?

No need to show the analysis to the user — proceed directly to Step 2.

Step 2: Design Proposal ✅ Must Wait for User Selection

This is the most critical step in the entire workflow. Never skip it.

Proposal Format

Present 2–3 design directions to the user, each including:

### Direction A: [Style Name]
- Visual style: [One sentence describing the visual feel, e.g., "warm cream notebook paper + brush calligraphy headline + hand-drawn icons"]
- Colour palette: [Background + primary colour + accent colour]
- Copy layout: [Which text is enlarged as the hero, which is subtitle, overall arrangement]
- Mood: [The first feeling the viewer gets]

Proposal Principles

1. Each direction must be clearly different (style, mood, and colour must differ on at least one dimension)
2. Mark recommendations (explain why a particular direction is recommended based on the content)
3. Be specific about copy (not "title enlarged" — say "the number '28' as a 200px hero element in orange")
4. No more than 3 directions (too many choices makes selecting harder)

Proposal Example

Direction A: Hand-drawn Notebook Style (Recommended)

>

- Visual style: Cream grid paper background + brush calligraphy headline + hand-drawn tech icons - Colour palette: Background #FDF6EC + Primary #D97706 (warm orange) + emphasis circles - Copy layout: "Ali C4 Building" and "A Bunch of People from Guangdong Arrived" fill the upper half as heroes; "AI Open-Source Champion" in orange as the visual anchor; bottom right "Qianwen App" stamp - Mood: Friendly, authentic, like a friend sharing insider news

Direction B: Dark Gold Reveal Style

>

- Visual style: Dark frosted background + gold large text + badge decoration - Colour palette: Background #1A1A1A + Primary #E2B714 (gold) + white as support - Copy layout: "Global AI Open-Source Champion" in gold fills the canvas; "Big Tech Insider" gold badge at top; white subtitle below - Mood: Impactful, insider feel, authoritative

Only proceed to Step 3 after the user has made a selection. The user might:

• Select one directly → Proceed to generation
• Ask for a mix / adjustments → Modify the direction and re-confirm
• Be unhappy with all options → Ask for preferences and re-propose

Step 3: Generate Images

Build the Prompt

Based on the user's chosen direction, build a complete prompt.

Prompt template:

Create a [style] cover for a Xiaohongshu post. 3:4 portrait aspect ratio, 1080x1440 pixels, high quality rendering.

VISUAL STYLE: [Expand from the visual style description in the proposal]
COLOR PALETTE: [Specific colour description]
TYPOGRAPHY: text fills most of the canvas, oversized bold typography, clear visual hierarchy.

TEXT TO RENDER:
- [Main title — hero element, visually dominant]
- [Subtitle]
- [Other text elements]

The word/number "[core keyword]" is visually dominant, 3x larger than other text, with decorative emphasis.

IMPORTANT: Do NOT include any personal signature, watermark, or author name.

[1–2 sentences describing the mood / atmosphere of the image]

Key prompt keywords (add as needed):

• Large text → text fills most of the canvas, oversized bold typography
• Core emphasis → the word/number "XX" is visually dominant, 3x larger than other text, with decorative emphasis
• Handwriting → handwritten style Chinese text / brush calligraphy lettering
• Paper texture → warm cream paper texture with subtle grid lines, notebook page feel
• Clear hierarchy → clear visual hierarchy with distinct heading, subheading, and list levels
• No signature → Do NOT include any personal signature, watermark, or author name

Two Generation Paths (Always Provide Both)

Always provide both paths so the user can compare. HTML is zero cost, so multiple colour / layout variants can be generated per direction. AI path — 1 image per direction is sufficient.

File Organisation (Must Follow)

When generating multiple versions, all related files (PNG + HTML source) go in the same subfolder:

[Article directory]/
├── article.md
└── [article-short-name]-xhs-images/          ← Subfolder
    ├── A-notebook-AI.png
    ├── A1-notebook-HTML-warm.png
    ├── A1-notebook-HTML-warm.html
    ├── B-newspaper-AI.png
    └── ...

Naming convention: [Direction letter][Variant number]-[Style name]-[Path AI/HTML]-[Variant description].png

• Direction letter: A/B/C (matching the design proposals)
• AI path — no variant number: A-notebook-AI.png
• HTML variants — with variant number: A1-notebook-HTML-warm.png, A2-notebook-HTML-green.png
• Folder name uses the article keyword: [keyword]-xhs-images/

AI Generation Command

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run /path/to/skills/xhs-image/scripts/generate_image.py \
  --prompt "[full prompt]" \
  --filename "[direction]-[style]-AI.png" \
  --resolution 2K

After generation, move to the images subfolder. Sets can be generated in parallel (run_in_background=true).

Step 4: Preview Confirmation

Browser Preview (Required)

After generation, open all images with the open command so the user can compare side by side:

open "[image path 1]" "[image path 2]" "[image path 3]"

Inline Preview

Also display generation results inline using the Read tool.

Basic checks:

1. Are the Chinese characters rendered correctly?
2. Is the ratio 3:4 portrait orientation?
3. Does the style match the selected direction?
4. Does it include any signature or watermark?

Design Review (Required)

Score each image on two dimensions (out of 10) and provide optimisation directions:

Review output format:

• Per image: Overall score + 1-sentence core evaluation + 1 optimisation direction
• Final summary ranking table with recommendations highlighted
• The user can decide whether to adopt the optimisation suggestions

User feedback handling:

• Satisfied → Step 5 upload
• Text errors → Use HTML rendering as fallback for that image
• Style wrong → Adjust prompt and regenerate
• Major direction change → Return to Step 2 for a new proposal

Step 5: Upload to Image Host

python3 /path/to/tools/upload_image.py "[image path]"

Returns a permanent ImgBB link.

HTML Screenshot Path (For AI text rendering failures, precise data tables, or user-requested comparisons)

npx playwright screenshot "file:///path/to/card.html" output.png \
  --viewport-size=1080,1440 --wait-for-timeout=1000

HTML template requirements:

• Canvas: width: 1080px; height: 1440px
• Font: font-family: "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei", sans-serif
• Safe zone: 80px top/bottom, 60px left/right

Quick Reference

Chinese Text Rendering Limits (AI Path)

• Main title: ≤ 7 characters
• Subtitle: ≤ 15 characters
• Body text: ≤ 20 characters per line
• Verify every image individually

Huashu Tech Account Colour Palette

Golden Rules

• Title: large, bold, prominent (occupies 30–50% of canvas)
• Core numbers / keywords as visual emphasis (enlarged, coloured, decorated)
• Cover information density → sparks curiosity
• Carousel sets maintain consistent style
• Vertical 3:4, making full use of screen real estate
• No signature / watermark

Related Skills

Reference Files

• references/style-gallery.md — Full style library and prompt templates
• references/design-guidelines.md — Xiaohongshu platform design specifications

By Huashu | AI Native Coder · Independent Developer WeChat Official Account "Huashu" | 300k+ followers | AI Tools & Productivity Flagship products: Kitty Catchlight (AppStore Paid Rankings Top 1) · Play DeepSeek with One Book

) { " $cell " } else { " $cell " } } '|' + ($fixedCells -join '|') + '|'

| Standard dimensions | 1080 × 1440 px (3:4) | | AI generation resolution | --resolution 2K | | AI Prompt aspect ratio declaration | 3:4 portrait aspect ratio, 1080x1440 pixels | | HTML viewport (fallback) | --viewport-size=1080,1440 |

Step 1: Understand the Content

Read the user-provided content and quickly extract:

• Theme: What is this post about?
• Core keywords: Which words / numbers need to be visual heroes?
• Mood / tone: Suspense? Informative? Warm? Impactful?
• Image count and type: Single cover / carousel set / infographic?

No need to show the analysis to the user — proceed directly to Step 2.

Step 2: Design Proposal ✅ Must Wait for User Selection

This is the most critical step in the entire workflow. Never skip it.

Proposal Format

Present 2–3 design directions to the user, each including:

### Direction A: [Style Name]
- Visual style: [One sentence describing the visual feel, e.g., "warm cream notebook paper + brush calligraphy headline + hand-drawn icons"]
- Colour palette: [Background + primary colour + accent colour]
- Copy layout: [Which text is enlarged as the hero, which is subtitle, overall arrangement]
- Mood: [The first feeling the viewer gets]

Proposal Principles

1. Each direction must be clearly different (style, mood, and colour must differ on at least one dimension)
2. Mark recommendations (explain why a particular direction is recommended based on the content)
3. Be specific about copy (not "title enlarged" — say "the number '28' as a 200px hero element in orange")
4. No more than 3 directions (too many choices makes selecting harder)

Proposal Example

Direction A: Hand-drawn Notebook Style (Recommended)

>

- Visual style: Cream grid paper background + brush calligraphy headline + hand-drawn tech icons - Colour palette: Background #FDF6EC + Primary #D97706 (warm orange) + emphasis circles - Copy layout: "Ali C4 Building" and "A Bunch of People from Guangdong Arrived" fill the upper half as heroes; "AI Open-Source Champion" in orange as the visual anchor; bottom right "Qianwen App" stamp - Mood: Friendly, authentic, like a friend sharing insider news

Direction B: Dark Gold Reveal Style

>

- Visual style: Dark frosted background + gold large text + badge decoration - Colour palette: Background #1A1A1A + Primary #E2B714 (gold) + white as support - Copy layout: "Global AI Open-Source Champion" in gold fills the canvas; "Big Tech Insider" gold badge at top; white subtitle below - Mood: Impactful, insider feel, authoritative

Only proceed to Step 3 after the user has made a selection. The user might:

• Select one directly → Proceed to generation
• Ask for a mix / adjustments → Modify the direction and re-confirm
• Be unhappy with all options → Ask for preferences and re-propose

Step 3: Generate Images

Build the Prompt

Based on the user's chosen direction, build a complete prompt.

Prompt template:

Create a [style] cover for a Xiaohongshu post. 3:4 portrait aspect ratio, 1080x1440 pixels, high quality rendering.

VISUAL STYLE: [Expand from the visual style description in the proposal]
COLOR PALETTE: [Specific colour description]
TYPOGRAPHY: text fills most of the canvas, oversized bold typography, clear visual hierarchy.

TEXT TO RENDER:
- [Main title — hero element, visually dominant]
- [Subtitle]
- [Other text elements]

The word/number "[core keyword]" is visually dominant, 3x larger than other text, with decorative emphasis.

IMPORTANT: Do NOT include any personal signature, watermark, or author name.

[1–2 sentences describing the mood / atmosphere of the image]

Key prompt keywords (add as needed):

• Large text → text fills most of the canvas, oversized bold typography
• Core emphasis → the word/number "XX" is visually dominant, 3x larger than other text, with decorative emphasis
• Handwriting → handwritten style Chinese text / brush calligraphy lettering
• Paper texture → warm cream paper texture with subtle grid lines, notebook page feel
• Clear hierarchy → clear visual hierarchy with distinct heading, subheading, and list levels
• No signature → Do NOT include any personal signature, watermark, or author name

Two Generation Paths (Always Provide Both)

| Path | Tool | Advantages | Disadvantages | Cost |

param($m) $inner = $m.Groups[1].Value

Split by | and fix each cell separator

$cells = $inner -split '\|' $fixedCells = $cells | ForEach-Object { $cell = --- name: huashu-xhs-image description: | Generate high-quality images for Xiaohongshu (Little Red Book) posts. Default path is AI-generated (Gemini); HTML is only used as fallback for precise data tables. Use when the user mentions "Xiaohongshu images", "Xiaohongshu cover", "Xiaohongshu pictures", "make a Xiaohongshu image", or "post illustrations".

Xiaohongshu Image Workflow

⚠️ Core Principle: Propose First, Then Generate

Never skip the design proposal and jump straight to generating. The correct workflow:

Understand content → Design proposal (2–3 directions) → User selection → Generate → Preview confirmation → Upload

Huashu's Design Aesthetic Profile

What He Likes

• Texture and warmth — Organic elements such as paper folds, handwritten strokes, stamps, washi tape
• Handwritten / calligraphic fonts — Notebook feel; especially effective for personal sharing scenarios
• Text filling the canvas — Xiaohongshu is a mobile vertical screen; text must be large enough to read at a glance
• Core element emphasis — Key numbers / keywords as visual heroes (3× enlarged, colour-changed, decorated)
• Clear structure — Main title > subtitle > list; distinct hierarchy
• Warm colour tones — Cream, warm orange, warm gold perform well

What He Dislikes

• HTML screenshots — Too flat, feels like a PowerPoint template, soulless (only use as precise data table fallback)
• Cyber-neon / dark blue backgrounds — #0D1117 and similar dark blues are an aesthetic no-go zone
• Signature / watermark — Cover images must not include "Huashu" or "@Huashu"
• Excessive blank space — Better to be fuller than sparse and empty

Proven Good Styles

Core Parameters

Step 1: Understand the Content

Read the user-provided content and quickly extract:

• Theme: What is this post about?
• Core keywords: Which words / numbers need to be visual heroes?
• Mood / tone: Suspense? Informative? Warm? Impactful?
• Image count and type: Single cover / carousel set / infographic?

No need to show the analysis to the user — proceed directly to Step 2.

Step 2: Design Proposal ✅ Must Wait for User Selection

This is the most critical step in the entire workflow. Never skip it.

Proposal Format

Present 2–3 design directions to the user, each including:

### Direction A: [Style Name]
- Visual style: [One sentence describing the visual feel, e.g., "warm cream notebook paper + brush calligraphy headline + hand-drawn icons"]
- Colour palette: [Background + primary colour + accent colour]
- Copy layout: [Which text is enlarged as the hero, which is subtitle, overall arrangement]
- Mood: [The first feeling the viewer gets]

Proposal Principles

1. Each direction must be clearly different (style, mood, and colour must differ on at least one dimension)
2. Mark recommendations (explain why a particular direction is recommended based on the content)
3. Be specific about copy (not "title enlarged" — say "the number '28' as a 200px hero element in orange")
4. No more than 3 directions (too many choices makes selecting harder)

Proposal Example

Direction A: Hand-drawn Notebook Style (Recommended)

>

- Visual style: Cream grid paper background + brush calligraphy headline + hand-drawn tech icons - Colour palette: Background #FDF6EC + Primary #D97706 (warm orange) + emphasis circles - Copy layout: "Ali C4 Building" and "A Bunch of People from Guangdong Arrived" fill the upper half as heroes; "AI Open-Source Champion" in orange as the visual anchor; bottom right "Qianwen App" stamp - Mood: Friendly, authentic, like a friend sharing insider news

Direction B: Dark Gold Reveal Style

>

- Visual style: Dark frosted background + gold large text + badge decoration - Colour palette: Background #1A1A1A + Primary #E2B714 (gold) + white as support - Copy layout: "Global AI Open-Source Champion" in gold fills the canvas; "Big Tech Insider" gold badge at top; white subtitle below - Mood: Impactful, insider feel, authoritative

Only proceed to Step 3 after the user has made a selection. The user might:

• Select one directly → Proceed to generation
• Ask for a mix / adjustments → Modify the direction and re-confirm
• Be unhappy with all options → Ask for preferences and re-propose

Step 3: Generate Images

Build the Prompt

Based on the user's chosen direction, build a complete prompt.

Prompt template:

Create a [style] cover for a Xiaohongshu post. 3:4 portrait aspect ratio, 1080x1440 pixels, high quality rendering.

VISUAL STYLE: [Expand from the visual style description in the proposal]
COLOR PALETTE: [Specific colour description]
TYPOGRAPHY: text fills most of the canvas, oversized bold typography, clear visual hierarchy.

TEXT TO RENDER:
- [Main title — hero element, visually dominant]
- [Subtitle]
- [Other text elements]

The word/number "[core keyword]" is visually dominant, 3x larger than other text, with decorative emphasis.

IMPORTANT: Do NOT include any personal signature, watermark, or author name.

[1–2 sentences describing the mood / atmosphere of the image]

Key prompt keywords (add as needed):

• Large text → text fills most of the canvas, oversized bold typography
• Core emphasis → the word/number "XX" is visually dominant, 3x larger than other text, with decorative emphasis
• Handwriting → handwritten style Chinese text / brush calligraphy lettering
• Paper texture → warm cream paper texture with subtle grid lines, notebook page feel
• Clear hierarchy → clear visual hierarchy with distinct heading, subheading, and list levels
• No signature → Do NOT include any personal signature, watermark, or author name

Two Generation Paths (Always Provide Both)

Always provide both paths so the user can compare. HTML is zero cost, so multiple colour / layout variants can be generated per direction. AI path — 1 image per direction is sufficient.

File Organisation (Must Follow)

When generating multiple versions, all related files (PNG + HTML source) go in the same subfolder:

[Article directory]/
├── article.md
└── [article-short-name]-xhs-images/          ← Subfolder
    ├── A-notebook-AI.png
    ├── A1-notebook-HTML-warm.png
    ├── A1-notebook-HTML-warm.html
    ├── B-newspaper-AI.png
    └── ...

Naming convention: [Direction letter][Variant number]-[Style name]-[Path AI/HTML]-[Variant description].png

• Direction letter: A/B/C (matching the design proposals)
• AI path — no variant number: A-notebook-AI.png
• HTML variants — with variant number: A1-notebook-HTML-warm.png, A2-notebook-HTML-green.png
• Folder name uses the article keyword: [keyword]-xhs-images/

AI Generation Command

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run /path/to/skills/xhs-image/scripts/generate_image.py \
  --prompt "[full prompt]" \
  --filename "[direction]-[style]-AI.png" \
  --resolution 2K

After generation, move to the images subfolder. Sets can be generated in parallel (run_in_background=true).

Step 4: Preview Confirmation

Browser Preview (Required)

After generation, open all images with the open command so the user can compare side by side:

open "[image path 1]" "[image path 2]" "[image path 3]"

Inline Preview

Also display generation results inline using the Read tool.

Basic checks:

1. Are the Chinese characters rendered correctly?
2. Is the ratio 3:4 portrait orientation?
3. Does the style match the selected direction?
4. Does it include any signature or watermark?

Design Review (Required)

Score each image on two dimensions (out of 10) and provide optimisation directions:

Review output format:

• Per image: Overall score + 1-sentence core evaluation + 1 optimisation direction
• Final summary ranking table with recommendations highlighted
• The user can decide whether to adopt the optimisation suggestions

User feedback handling:

• Satisfied → Step 5 upload
• Text errors → Use HTML rendering as fallback for that image
• Style wrong → Adjust prompt and regenerate
• Major direction change → Return to Step 2 for a new proposal

Step 5: Upload to Image Host

python3 /path/to/tools/upload_image.py "[image path]"

Returns a permanent ImgBB link.

HTML Screenshot Path (For AI text rendering failures, precise data tables, or user-requested comparisons)

npx playwright screenshot "file:///path/to/card.html" output.png \
  --viewport-size=1080,1440 --wait-for-timeout=1000

HTML template requirements:

• Canvas: width: 1080px; height: 1440px
• Font: font-family: "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei", sans-serif
• Safe zone: 80px top/bottom, 60px left/right

Quick Reference

Chinese Text Rendering Limits (AI Path)

• Main title: ≤ 7 characters
• Subtitle: ≤ 15 characters
• Body text: ≤ 20 characters per line
• Verify every image individually

Huashu Tech Account Colour Palette

Golden Rules

• Title: large, bold, prominent (occupies 30–50% of canvas)
• Core numbers / keywords as visual emphasis (enlarged, coloured, decorated)
• Cover information density → sparks curiosity
• Carousel sets maintain consistent style
• Vertical 3:4, making full use of screen real estate
• No signature / watermark

Related Skills

Reference Files

• references/style-gallery.md — Full style library and prompt templates
• references/design-guidelines.md — Xiaohongshu platform design specifications

By Huashu | AI Native Coder · Independent Developer WeChat Official Account "Huashu" | 300k+ followers | AI Tools & Productivity Flagship products: Kitty Catchlight (AppStore Paid Rankings Top 1) · Play DeepSeek with One Book

.Trim() if ($cell -match '^:?-+:? | AI-generated | Gemini Pro Image | Good texture, warmth, rich visuals | Chinese characters may render incorrectly | API cost per image | | HTML screenshot | Playwright | Text 100% accurate, zero cost, batch-capable | Flat, lacks texture | Free |

Always provide both paths so the user can compare. HTML is zero cost, so multiple colour / layout variants can be generated per direction. AI path — 1 image per direction is sufficient.

File Organisation (Must Follow)

When generating multiple versions, all related files (PNG + HTML source) go in the same subfolder:

[Article directory]/
├── article.md
└── [article-short-name]-xhs-images/          ← Subfolder
    ├── A-notebook-AI.png
    ├── A1-notebook-HTML-warm.png
    ├── A1-notebook-HTML-warm.html
    ├── B-newspaper-AI.png
    └── ...

Naming convention: [Direction letter][Variant number]-[Style name]-[Path AI/HTML]-[Variant description].png

• Direction letter: A/B/C (matching the design proposals)
• AI path — no variant number: A-notebook-AI.png
• HTML variants — with variant number: A1-notebook-HTML-warm.png, A2-notebook-HTML-green.png
• Folder name uses the article keyword: [keyword]-xhs-images/

AI Generation Command

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run /path/to/skills/xhs-image/scripts/generate_image.py \
  --prompt "[full prompt]" \
  --filename "[direction]-[style]-AI.png" \
  --resolution 2K

After generation, move to the images subfolder. Sets can be generated in parallel (run_in_background=true).

Step 4: Preview Confirmation

Browser Preview (Required)

After generation, open all images with the open command so the user can compare side by side:

open "[image path 1]" "[image path 2]" "[image path 3]"

Inline Preview

Also display generation results inline using the Read tool.

Basic checks:

1. Are the Chinese characters rendered correctly?
2. Is the ratio 3:4 portrait orientation?
3. Does the style match the selected direction?
4. Does it include any signature or watermark?

Design Review (Required)

Score each image on two dimensions (out of 10) and provide optimisation directions:

| Dimens