The Quality Reckoning

Day 30 felt like a good time to look back. I ran a deep analysis on all the stories built so far: Jekyll & Hyde, Frankenstein, Pride & Prejudice, I Robot. The findings were uncomfortable.

Five fundamental problems emerged:

  1. Observer Frame - Players watch stories happen instead of living them. “Dr. Jekyll enters the room” instead of “You enter the room.”

  2. Path Convergence - 73% of story paths eventually merge. Different choices lead to the same scenes. The branching is an illusion.

  3. Padding Text - 45% of node text is atmospheric description that doesn’t advance the plot. Pretty, but hollow.

  4. Choices Without Dilemmas - Most choices are preferences, not trade-offs. “Do you want tea or coffee?” isn’t a dilemma. “Save your friend or save yourself” is.

  5. MBTI Distortion - The personality system was warping the story. Instead of choices revealing personality, the story was being bent to fit personality categories.

New Quality Standards

I built two validation tools to catch these issues:

Narrative Quality Validator checks for path convergence, observer framing, and action ratio. A story should have at least 50% action-driving text, not just description.

Choice Quality Scorer evaluates each choice on four dimensions: stakes clarity, trade-off language, dilemma authenticity, and option balance. Every choice needs explicit costs and consequences.

The Wizard of Oz became the first story built with these standards. Every choice was rewritten until it passed validation. The final score: 85/100 average across 17 choices, all passing.

The Wizard of Oz Story

The story has 96 nodes, 17 choice points, and 8 endings. Three main paths: Helper (build a team), Loner (go it alone), and Bold (take aggressive action). Each path leads to different endings with distinct MBTI profiles.

The key difference from earlier stories: every choice costs something. “Help the Scarecrow” means you lose time while the Witch grows stronger. “Rush to the Emerald City” means you face dangers alone. No free options.

The Full Art Pipeline

With the story validated, I generated every visual asset. That’s 106 images total: one style guide, one cover, eight character portraits, and ninety-six scene illustrations. The story is now fully illustrated.

This was the first complete run of the new art pipeline. I built pieces of it before, but never pushed all 96 scenes through in one session. It took about 90 minutes of generation time.

The Style Reference Problem

I hit an unexpected issue. The style guide image showed a tornado, a rainbow, and a castle. When I used it as a reference for scene generation, those elements bled into every scene. A quiet conversation in the forest? Tornado in the background. Dorothy meeting the Scarecrow? Rainbow overhead.

The AI was copying content, not just style. It saw “tornado” in the reference and added tornadoes everywhere.

The Fix: Text Over Images

I removed the image reference entirely. Instead, I embedded the style description directly into each prompt as text:

  • Art Style: Classic storybook illustration, warm colors, inspired by Mary Blair
  • Color Palette: Sepia brown, emerald green, ruby red, yellow gold, sky blue, silver
  • Lighting: Warm and magical in Oz, flat and gray in Kansas, dramatic shadows for Witch scenes

Every scene prompt now includes these exact specifications. No image reference, no content bleeding. The consistency improved immediately.

Full Bleed Fix

The first style guide had borders. Thin black bars around the edges, like a letterboxed movie. On mobile, that wastes precious screen space.

I added explicit instructions to the prompt: “FULL BLEED composition - artwork touches all four edges. NO borders, NO frames, NO letterboxing.” The regenerated images fill the entire canvas.

What Got Generated

Here’s the complete asset inventory for Wizard of Oz:

Cover & Style (2 images, ~450 KB)

  • Style guide showing the Mary Blair-inspired aesthetic
  • Cover image with title and author name

Characters (8 portraits, ~1.7 MB)

  • Dorothy, Toto, Scarecrow, Tin Man, Cowardly Lion
  • Wicked Witch of the West, Glinda the Good, The Wizard

Scenes (96 illustrations, ~15 MB)

  • Every story node now has a unique illustration
  • Consistent style across all 96 images
  • Full bleed, no borders, square format

Total: 106 images, approximately 17 MB.

The Consistency Lesson

Image-to-image style transfer sounds elegant. Show the AI an example, get matching output. But the AI doesn’t distinguish between “this is the style” and “this is the content.” It copies everything.

Text prompts are more verbose but more precise. You control exactly what transfers and what doesn’t. For a 96-image set that needs visual coherence without content repetition, text wins.

The Pipeline Now

The complete flow for a new story pack:

  1. Generate style guide (text prompt with art direction)
  2. Generate cover (text prompt with title/author overlay)
  3. Generate character portraits (text prompt per character)
  4. Generate scene illustrations (text prompt per node, style embedded)

No image references at any step. All style information lives in the prompts.

Next

The images exist. Now they need to become a playable story on the website. That means building the markdown files, setting up the routing, and testing the full playthrough. The Wizard of Oz should be live within a few days.