wonder-engine

Serious ideas, turned into story-worlds you can wander through.

Wonder Engine turns papers, research directions, school topics, technical trends, and private fascinations into strange, clever illustrated books. It treats an idea not as a lesson to summarize, but as a world to build: concepts become creatures, maps, rituals, rules, weather, craft, food, games, machines when they truly fit, jokes, and trouble.

The process begins with a creative interview, then moves through source digestion, story architecture, continuity anchors, one-spread-at-a-time image generation, careful text layout, and a polished PDF. The aim is a book with real narrative gravity: odd and science-flavored, visually dense but readable, playful enough for children and layered enough for adults.

What It Makes

Example spread: English story text is typeset over native quiet space in the illustration.

A finished Wonder Engine book can include:

• portrait front cover
• opening character-introduction spread with portraits, avatars, prop cards, or cast vignettes that anchor character continuity
• opening world-introduction spread with a map, object atlas, concept diagram, ritual chart, garden plan, recipe board, machine, or civic system that anchors world logic
• 20 double-page story spreads by default
• optional back matter explaining the real idea under the adventure
• optional portrait back cover with a short blurb, source note, or production credit
• final PDF with manually typeset story text and audited native labels where useful

The result should feel like an adventure, not a textbook or a captioned storyboard: characters argue, test, misunderstand, repair, escape, and discover. Abstract ideas become visible things: maps, rituals, creatures, tools, diagrams, hazards, rules, gardens, kitchens, weather signs, games, or machines when the story really needs them.

Example Book

Maren & Knapp and the Blue Doors of Norrvik: a complete Fjord-story PDF produced with the workflow.

Open the complete PDF

How It Works

1. Interview
   A natural creative interview that uses nudges, contrasts, and occasional options to help the human discover what they mean: source material, audience, feeling, story mode, world flavor, culture/language, cast ecology, motifs, avoidances, length, and output.

2. Digest the Source
   The workflow extracts the real thesis, mechanisms, vocabulary, stakes, and limits, then maps them into story objects and visual metaphors.

3. Pitch the Book
   Imaginative title/tagline packages, lead/cast options, world, core metaphor table, chapter structure, and assumptions.

4. Optional Visual Probe
   One disposable concept image can test the world, cast ecology, material balance, density, and mood before the full story is planned. The probe is for feedback, not canon.

5. Continuity Bible
   The workflow locks character anchors, world rules, recurring objects, map logic, visual motifs, and reusable prompt snippets before main story images are generated.

6. Plan the Spreads
   A coherent chapter-by-chapter beat plan built from a plain causal story spine, where every spread has an inbound bridge, visible event, decision or consequence, and outbound bridge.

7. Write and Prompt
   Final spread text, image prompt, composition notes, text-space plan, speech bubble notes, and print safety notes.

8. Generate and Assemble
   Images are generated one by one, using the continuity bible and approved intro spreads as anchors. The image model creates the artwork, not the final readable manuscript. Story text, headings, cover titles, credits, and other exact wording are overlaid afterward as real PDF typography.

At every stage, Wonder Engine should also guide the human toward the next useful decision or review: what was just produced, what comes next, and what kind of feedback will move the book forward.

Image Generation

Wonder Engine is image-generator agnostic.

Use the native image generation available in the current environment, keeping the same spread prompts, one-image-at-a-time review loop, and PDF assembly rules.

If more than one image model or image-generation tool is available, ask the human which one they want before the first visual probe or production spread. Wonder Engine is model-agnostic, but its current prompt guidance, native-label policy, and layout-proof workflow have been tested on GPT Image 2.0, so GPT Image 2.0 is the recommended default unless the human chooses another model.

Artwork should be generated in the format it needs to occupy: portrait cover, portrait back cover, wide hero, or seamless landscape interior. Cropping is only a tiny safe trim, not a way to turn the wrong composition into the right one. If the ratio is wrong, regenerate or extend the image.

When a back cover is part of the book, it gets its own generated image and text area. It should not be skipped, left as a placeholder, or made by recycling the front cover unless the human specifically asks for a wraparound design.

Text Is Not Image-Generated

Wonder Engine does not ask the image model to render the final manuscript. The art is generated first with intentional quiet space, then the exact text is placed over the approved images as editable PDF typography.

That keeps the book readable, correctable, and easier to translate. It also means titles, chapter headings, story text, cover credits, important speech bubbles, and any wording that must be exact are not trusted to the image model.

There is one limited exception: diagram-like anchor pages can use native image text for short integrated labels, names, badges, map marks, arrows, and tiny callouts when those labels need to live inside the art. This is mainly for pages such as cast introductions, world maps, object atlases, or concept diagrams. Those labels should be brief and reader-helpful, and they must be checked against a locked label sheet before approval.

Multiple Languages

Wonder Engine can shape a book for different languages, scripts, and reading contexts. Language is part of the creative brief: it affects interview questions, source handling, typography, line length, writing direction, cultural signals, image prompts, and final PDF layout.

The image model still creates artwork, not final manuscript text. For localized or multilingual books, the workflow chooses fonts with the right script support and overlays the exact approved text after the illustration is generated.

Ukrainian sample: wind, seeds, and making invisible movement visible. Artwork generated first; Ukrainian text overlaid afterward.

Japanese sample: paper fibers, load, and material testing without a machinery-first metaphor. Artwork generated first; vertical Japanese text overlaid afterward.

The Layout Idea

Wonder Engine does not simply paste text over a white rectangle.

It plans native quiet space inside the illustration before image generation: mist, sky, water, paper, wall, snow, empty table, map margin, or another calm area that belongs to the scene. The spread text and artwork are designed together, so the quiet area matches the actual amount of prose.

Flat white cards are not the fallback. If a story spread needs readable body text, the image should naturally provide the lighter area for it. If the generated art does not, the right fix is usually to regenerate, split the text into better islands, or revise the layout contract.

The current PDF helper includes a negative-space fitting pass:

python3 wonder-engine/scripts/new_manifest.py \
  --title "Working Title" \
  --subtitle "A Quirky Tagline" \
  --author "by Author Name" \
  --spreads 20 \
  --back-cover \
  --output build/book-manifest.json

python3 wonder-engine/scripts/fit_text_to_negative_space.py \
  build/book-manifest.json \
  --output build/book-manifest-fitted.json

python3 wonder-engine/scripts/assemble_picture_book_pdf.py \
  build/book-manifest-fitted.json \
  --output build/book-layout-proof.pdf \
  --preview-dir build/previews

The fitter measures pale, low-detail areas in the actual generated art and writes shaped text rows into the manifest. It is a proofing tool, not a substitute for taste: if a spread still feels awkward, revise the prose, regenerate the art, split the text, or move the title/body composition.

The first assembled PDF is a layout proof, not the finished book. Wonder Engine renders page previews, checks the actual usable quiet space in each image, adjusts text boxes and line shapes, and only calls the PDF final after every page looks intentionally typeset.

Chapter-start spreads get extra layout scrutiny: the chapter label, chapter title, and body text are treated as one stacked composition so the title does not float away from the text area or collide with the body copy.

Typography

Wonder Engine asks about typography during the interview for full PDFs. The rule is simple: use a real pairing, not one generic fallback font. Large titles and body text should usually be different fonts with clear roles.

Good pairings for this kind of book:

Mood	Title Font	Body Font	Why It Works
Current sample spread	Cochin	Iowan Old Style	The pairing used in the current proof spread: elegant, bookish titles with a warm readable old-style body.
Curious, inky, old-machine	Fraunces	Literata	Fraunces gives titles a playful old-style weirdness; Literata is built for sustained reading and keeps the page calm.
Handmade workshop	Eczar	Alegreya	Eczar feels sturdy, hand-cut, and storybook-strange; Alegreya keeps the reading warm and literary. Good for tools, markets, diagrams, and comic engineering.
Elegant cabinet of curiosities	Cormorant Garamond	Source Serif 4	Cormorant is ornamental and dramatic at large sizes; Source Serif 4 gives the manuscript steadier rhythm.
Bold theatrical adventure	Playfair Display	Libre Baskerville	Strong, high-contrast titles with a familiar, readable bookish body. Best when the art has a little drama.
Multilingual or translation-first	Noto Serif Display	Noto Sans	A practical starting point when language coverage matters more than decorative personality. Swap in script-specific Noto families as needed.

These are examples, not limits. Choose other pairings when the language, culture, trim size, or book personality calls for it.

Illustration Direction

The current skill focuses on whimsical, dense, science-flavored adventure illustration: rich scenes, visible systems, expressive main characters, clean negative space, and simplified background figures. Machinery is available when it belongs, but the same workflow can make science visible through paper, weather, gardens, kitchens, music, animals, rituals, maps, games, and everyday experiments.

More illustration style presets are planned for the future. The goal is to support different visual traditions, cultures, languages, moods, and production needs without copying any existing book, artist, character, or protected style.

What Is Codified Today

• front cover generation and manual cover typography
• distinctive title plus subtitle/tagline guidance
• opening character-introduction spread as a character continuity anchor
• opening world-introduction spread as a world-logic anchor
• continuity bible for characters, world rules, map logic, recurring objects, and prompt snippets
• native-text label policy for character/world anchor spreads, with locked label sheets and proofreading
• optional back matter
• optional back cover
• target-ratio generation for cover, hero, spread, and back-cover art
• one-spread-at-a-time image generation and approval
• natural text-space planning before image generation, with no generic white-card rescue
• show-don't-tell manuscript rules
• agent-owned causal story-spine and cold-reader checks
• spread-to-spread "because/but/therefore" continuity gates
• whimsy sense audit so clever phrases stay meaningful
• story-density guidance so spreads become full scenes, not two-sentence captions
• source-to-story metaphor mapping
• culture and language intake
• typography intake and font pairing guidance
• negative-space-aware PDF assembly
• page-by-page layout proofing before final PDF export
• chapter-start title/body collision checks
• forward-guidance behavior after each major stage
• exact text typesetting over approved images

Files

• wonder-engine/SKILL.md - core skill instructions
• wonder-engine/references/interview-options.md - option-led interview prompts
• wonder-engine/references/source-processing.md - source digestion and metaphor mapping
• wonder-engine/references/continuity.md - character/world continuity bible and anchor-spread rules
• wonder-engine/references/story-standards.md - story, image, typography, and PDF quality rules
• wonder-engine/scripts/new_manifest.py - starter PDF manifest
• wonder-engine/scripts/fit_text_to_negative_space.py - text-shape fitting helper
• wonder-engine/scripts/assemble_picture_book_pdf.py - PDF proof/final assembly helper

References Used for Typography

• Current proof spread fonts: Cochin for title/chapter text and Iowan Old Style for body text.
• Google Fonts: Fraunces, Literata, Eczar, Alegreya, Cormorant Garamond, Source Serif 4, Playfair Display, and Libre Baskerville
• Noto documentation for multilingual font coverage and licensing notes
• Typewolf on headline/body Google Font pairings for the principle of pairing a large-display face with a body face built for reading