Arcanea Book Cover

Arcanea Book Cover Design Skill

Generate premium book covers by thinking deeply about the book first, then translating that thinking into a precision image prompt. NB2 is default. NB Pro only when user specifies "premium" or "hero".

The 5-Phase Method

Do NOT skip phases. The thinking is the skill. The generation is the reward.

Phase 1: READ THE BOOK (or its architecture)

Before writing any prompt, you must understand:

1. Central tension — What does the protagonist fight against? What internal conflict mirrors the external?
2. Emotional register — Is this grief-heavy? Wonder-heavy? Dread-heavy? A dark comedy? Bittersweet? Name it precisely.
3. Iconic image — What ONE visual moment defines the book? The image a reader would tattoo on themselves.
4. Genre expectations vs subversion — What do readers of this genre expect? What should you give them, and what should you withhold?
5. Color emotional key — What color (singular) carries the book's feeling?

If the book has a book.yaml or outline, read them. If the book has existing chapters, read at least one opening and one pivotal scene.

Phase 2: PULL BEST PRACTICES

Top book cover designers operate on these principles. Keep them active during prompt design:

Chip Kidd's Rules:

• One idea, one image, one emotion
• The cover is a question, not an answer — it should make readers curious, not satisfied
• Text and image are one unit — typography is 50% of the cover
• Don't illustrate the plot. Illustrate the feeling.

Peter Mendelsund's Rules:

• Covers are "visual metaphors, not visual illustrations"
• Abstraction > literalism when the book is literary
• The reader should finish the book and look at the cover and think "of course"
• White space is content

Coralie Bickford-Smith's Rules (Penguin Classics):

• Pattern + symbol > scene (for literary)
• Limited palette (2-4 colors max)
• Typography must be the hero on pattern-based covers
• Consistency within a series, uniqueness between series

Genre-specific heuristics:

Universal principles:

• Title first: Typography must be readable at thumbnail size (Amazon grid)
• Composition rule of thirds: Focal point on intersection, not dead center
• Depth layering: Foreground / mid-ground / background — create dimension
• One hero color: Everything else supports it
• Resist clutter: If it's not essential, delete it
• Portrait ratio: 2:3 is the book cover standard (1200x1800, 800x1200, etc.)

Phase 3: COMPOSE THE PROMPT

The prompt has a specific structure. Follow it exactly:

Generate an image: [Genre descriptor] book cover for [TITLE]. Portrait orientation, 2:3 ratio.

SCENE: [One paragraph describing the single iconic image. Concrete. Specific. No abstractions.]

COMPOSITION: [Focal point location, depth layers, what's foreground/mid/background]

LIGHTING: [Source, direction, mood. Cinematic lighting terminology.]

PALETTE: [Hero color + 2-4 supporting colors as hex codes or named precisely]

TYPOGRAPHY: [Title at top/bottom, font style: serif/sans/display, weight: bold/thin]

STYLE REFERENCES: [2-3 specific references. Real designers, movies, or artists.]

MOOD: [One sentence on the emotional register the cover should evoke]

TECHNICAL: 4K resolution, portrait ratio, book cover composition, cinematic depth of field, readable typography.

Phase 4: GENERATE

Model routing:

• DEFAULT: NB2 (gemini-3.1-flash-image-preview) — $0.02, portrait-native, best for covers
• PREMIUM (only when user says "premium", "hero", "final", or "launch"): NB Pro (nano-banana-pro-preview) — $0.04

API call template:

# Load key from MCP config or env
# export GEMINI_API_KEY="your-key-from-aistudio.google.com"

MODEL="gemini-3.1-flash-image-preview"  # NB2 default
# MODEL="nano-banana-pro-preview"        # NB Pro (premium only)

curl -s "https://generativelanguage.googleapis.com/v1beta/models/$MODEL:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"PROMPT_HERE"}]}],"generationConfig":{"responseModalities":["TEXT","IMAGE"]}}' \
  | python3 -c "
import json,base64,sys
data=json.load(sys.stdin)
for p in data.get('candidates',[{}])[0].get('content',{}).get('parts',[]):
    if 'inlineData' in p:
        d=base64.b64decode(p['inlineData']['data'])
        out='apps/web/public/images/books/SLUG-cover.png'
        open(out,'wb').write(d)
        print(f'OK: {len(d)} bytes -> {out}')
    elif 'text' in p:
        print('Text:',p['text'][:200])
"

Save path convention: apps/web/public/images/books/[slug]-cover.png (Add -nb2 or -nbpro suffix if generating multiple for comparison.)

Phase 5: VALIDATE

Before declaring the cover done, check:

• Title is readable — Can you read it at 200px wide (thumbnail size)?
• Genre signal is clear — Would a reader in 1 second know the genre?
• One hero color dominates — Not competing palettes
• Focal point is intentional — Rule of thirds or centered deliberately
• No AI artifacts — Hands, text garbling, extra limbs, muddy areas
• Portrait ratio — Not square, not landscape
• Matches book tension — Would someone who read the book nod in recognition?
• Mood is precise — Not generic "epic" or "dark" — the SPECIFIC mood

If any fail, iterate. One re-generation is fine. Beyond that, revise the prompt.

Quality Tiers

• Draft cover — NB2, one generation, basic validation. Good enough for book-in-progress.
• Production cover — NB2, 2-3 generations, full validation, pick best. Default for published works.
• Premium cover — NB Pro, multiple angles, hero-quality. For launches, featured tier, or when user specifies.

Anti-Patterns (Never Do)

• Generic "fantasy book cover" prompts with no specificity
• Character faces unless specifically required (faces are the hardest AI generation)
• More than one human figure on the cover (crowd scenes fail)
• Trying to illustrate the entire plot in one image
• Using NB Pro by default (costs 2x with marginal benefit for most covers)
• Skipping Phase 1 (reading the book) to save time
• Bright neon palettes on literary books
• Embedded text longer than the title (taglines often fail in generation)

Integration with Arcanea Ecosystem

• Save covers to apps/web/public/images/books/[slug]-cover.png
• Update COVER_MAP in apps/web/app/books/drafts/[slug]/page.tsx
• Update COVER_MAP in apps/web/app/books/drafts/page.tsx
• Log the cover generation in the book's book.yaml under ai_transparency.models_used
• For featured tier books: also register in the book reader's metadata

Output Format

When using this skill, produce:

1. The thinking — A brief (150-250 word) analysis of the book's tension, emotion, iconic image, genre expectations
2. The prompt — The composed prompt with all 7 structural sections
3. The generation — Execute the curl command, save to correct path
4. The validation — Run the checklist, report pass/fail per item
5. The result — Read the generated image and confirm visually

"A cover is a haiku. One image, one emotion, one question that makes you turn the book over to read the back." — The Arcanea Book Cover Method