Zild: Echoes Under Two Moons - Cover Art

This novel is, at its core, an experiment in pushing AI beyond short-form bursts into the realm of sustained, coherent storytelling. By orchestrating dozens of segmented prompts, weaving in summaries, and stitching together multi-thousand-word chapters, I’ve tested how an AI co-author can maintain narrative momentum, character arcs, and thematic consistency over an 18-chapter canvas. Every journal entry, subplot seed, and visual prompt has been a deliberate probe into the model’s capacity for long-form cohesion and creative depth.

Along the way, I’ve navigated several key hurdles: the finite context window that forces chunked inputs and multi-pass expansions; the tendency of the model to repeat itself, requiring deduplication and manual polishing; and the challenge of preserving voice and subplot continuity as the draft grows. Image generation adds its own iterative loop—fixing hand shapes, fire intensity, and framing until the artwork matches the prose. Through careful tooling, bullet-point scaffolds, and human oversight, I’ve carved a path for AI to contribute meaningfully to an epic-length narrative, while acknowledging that collaboration still demands rigorous editorial guidance.

Behind the Scenes of Crafting Zild: A Deep Dive into the Creative Workflow

Published on DtRH – Down the Rabbit Hole
By Kevin “KBS” Bradley | April 26, 2025

Ever wondered how a sprawling multi-chapter sci-fi epic springs from a handful of prompts and an AI co-pilot? In this article, I’ll pull back the curtain on Zild, my 18-chapter novel set on an alien world that still echoes Earth’s legacy. From the first role-play seed to the final cover art, here’s every step, tool, and trick that turned a concept into a cohesive, richly detailed saga.

1 Planting the Seed: Defining “Spacy”

It all began with a single line of instruction:

You are Spacy, a bio-machine historian on planet Zild, charged with preserving humanity’s story.

Within seconds, the AI “woke up” in character, drafting a mission statement that felt both poetic and authoritative. I then peppered it with questions like “What led to Earth’s downfall?” and “Write your first journal entry,” locking in a consistent voice and world logic. Screenshots from this phase show the earliest sparks of that living archive persona (to be added). There was no particular plan at this point - I simply asked questions and allowed it to answer. I probed it in general directions I wanted, but let it do the a lot of the legwork.

2 Mapping the Journey: The 18-Chapter Arc

Rather than let the narrative wander, I sketched a high-level roadmap:

1. Chapters 1–3: The Last Days of Earth
2. Chapters 4–6: The Ark’s Launch and Voyage
3. Chapters 7–9: First Survival on Zild
4. Chapters 10–13: Building a New Home
5. Chapter 14+: Epilogue & Final Reflections

By slicing the story into these five acts, I kept each segment focused—and avoided the dreaded “context-window overload” that plagues long AI sessions. Without having a road to follow, things can fall apart pretty easily.

3 Assembling the Draft: From Fragments to z.txt

Over hours/days, I collected dozens of AI-generated snippets—journal entries, dialogue, environmental descriptions. Each chunk went into a master text file, affectionately dubbed z.txt (about 200 KB by the end after deduping). This single-file approach let me search for recurring themes, spot continuity issues, and ensure no subplot drifted off course.

4 Naming the Chapters: From Placeholders to Polished Titles

With raw drafts in hand, I asked the AI: “Suggest three titles for Chapter N.” For example, Chapter 1 yielded:

• Fading Horizons
• Echoes of a Dying World
• The Final Countdown

I blended or tweaked these until each title felt evocative, then baked them into my markdown front-matter. Now, every chapter file read like a mini-story in its own right. In many chapters these names have changed, but they served as a very brief explanation of said chapter, and helped with following the story as vaguely designed in the chapter outline phase.

5 Beating the Length Limits: Two-Pass Expansion

AI models struggle when forced to output thousands of words in one go. My solution:

• Pass A: “Expand Chapter N to at least 2,500 words. Stop at part 1.”
• Pass B: After I type “OK,” the model continues with part 2, hitting the target length.

I also prefaced each prompt with a brief summary of earlier chapters to maintain narrative continuity. The result: richly detailed chapters without token-limit blowups. This is easier said than done, but that explains the general process. If you notice your context window is something like 1500 words, and you want a 4500 word chapter - You should draft a chapter that is around 1500 words (no more), split it into three chunks, break down the important points of each chunk, then generate the chapter in 3 parts. At the end, consolidate them into one chapter.

6 Deduplication & Editorial Polishing

Large-scale expansion often introduces accidental repetition. I wrote a quick Python script to filter out duplicate lines:

import re, pathlib

def dedupe(text):
    seen, out = set(), []
    for line in text.splitlines():
        signature = line.lower().strip()
        if signature not in seen:
            seen.add(signature)
            out.append(line)
    return "".join(out)

Then, with regex-aided passes and a final manual polish, I ironed out awkward redundancies and sharpened the prose.

7 Seeding Subplots & Deepening Characters

Every chapter got a fresh prompt:

“Insert one new subplot seed; do not resolve it yet. Add deeper sensory detail and internal monologue.”

This kept the narrative tapestry dense but coherent. I cross-checked new threads against my master “Unresolved Plot Threads” list to avoid accidental overlap.

8 Building the Glossary & Appendix

To help readers navigate Zild’s technology, biology, and backstory, I maintained two living documents:

• glossary.md – definitions of unique terms
• appendix.md – tables of planetary data, selection algorithms, key specifications

After each chapter pass, I asked the AI: “Update glossary with any new terms. Add rows to the appendix if new tech appears.” The merged result became a robust reader resource. The Appendix has served as a way to follow sub plots that have been seeded, allowing me to resolve them when I see fit and not forget about them

9 Dropping Easter Eggs: Personal & External

There are quite a few in here. For instance, in Chapter 1’s finale, I replaced a placeholder quote with a poem decoded from a Base64 blob hidden in my bunny.js project on GitHub. This small insertion created a bridge between my code experiments and the novel’s lore.

10 Illustrating the Story: DALLE in the Loop

Each chapter demanded a striking header image (1200 × 600 px):

1. Select a passage that captures mood.
2. Craft a DALLE prompt (e.g., “feminine hand on viewport, tears floating, scorched Earth below”).
3. Iterate until details align: hand shape, fire intensity, framing.
4. Embed final PNG in chapter markdown with a lightbox viewer.

Screenshots from these iterations show how I nudged the AI to fix gender, lighting, and color balance.

11 Conquering Context Limits: Bullet-Point Summaries

When the later chapters plus all lore exceeded the model’s capacity, I distilled each fully written chapter into a 200–300-word bullet summary. Feeding those notes plus the raw new chapter prompt let the AI recombine threads while preserving crucial events and character arcs.

12 Putting It All Together: Project Structure

zild/
├─ drafts/           # individual chapter .md files
├─ notes/            # bullet-point summaries
├─ images/           # chapter header art
├─ glossary.md
├─ appendix.md
└─ z.txt             # monolithic master draft

Git version control kept every iteration safe and traceable under DTRHnet/zild on GitHub (private repo).

Additionally, I’ve been using Markdown Novel Template to compile the novel as I go into drafts in formats such as epub, pdf, md, txt, etc. This cool project allows you to do this all by running a simple shell script!! I haven’t been posting any drafts yet, but will eventually offer the book for download across all file formats.

13 The Toolbelt

• ChatGPT (o3 & 4o) – primary prose engine
• Python + regex – text deduplication & transformation
• VS Code – markdown editing, preview, Mermaid diagrams
• DALLE-3 – chapter illustrations
• Git – collaborative versioning
• Markdown Novel Template

14 Lessons Learned & Future Directions

• Chunking wins. Small, focused prompts avoid token overload.
• Note-taking matters. Bullet summaries preserve narrative without drowning in detail.
• Human oversight remains essential for voice consistency and polish.