The numbers

A natural-prose retelling of v1, same arc and same illustrations. 37% shorter in words, 20 minutes shorter in audio. The verse stayed only where it earned its keep — about a third of v1's line count.

The friends sat close on the cobblestones — closer than they had sat all week. No one spoke. A lantern across the square guttered and steadied.

Mango's wings were folded so tight they almost disappeared.

She spoke first.

"I didn't close the fuel." Her voice was low, stripped of every exclamation mark. "I knew the step. I've sung it back a hundred times. But my hands were already reaching for the spin, and my mouth said the words and my hands did not do them."

The cast — and their voices

April

Sarah

Rainbow koala teacher

Mango

Jessica

Fastest fruit bat

Fern

Lily

Careful tree frog

Riddle

Laura

Questioning parrot

Dusty

Roger

Steady wombat

Cypher

Bill

April's old teacher (Ch. 9) — older tiger, spins poi

Plus George (Narrator). Cypher replaces Laurel from v1 — he is based on the inspirer's real teacher; he spins poi on chains (not staff), and his voice is Bill — a wise mature register.

Cost breakdown

v2 reused all 98 v1 illustrations + the cover, so image-gen cost was almost nothing — just one new Cypher portrait (~$0.07) and zero re-renders. TTS dropped from 490 to 400 segments. The big metered cost line stays ElevenLabs.

The v2 savings. v2 didn't run arc fan-out, judging, spread, character generation, factor generation, cover render, or 5 of the 6 portrait renders — all of those were inherited from v1. The only fresh model calls were the 10 chapter weaves (with the new style guide + new safety steps), 10 audio segmentations, and 1 new Cypher portrait. ElevenLabs was the only line that scaled with the new prose, and it dropped from 490 to 400 segments — call-and-response chapters used to require a tag-switch per couplet; natural prose lets the narrator carry longer stretches before a voice swaps in.

Timing

v2 was fast because most stages were reused from v1. Active runtime for the new work: about 10 minutes.

What changed in v2

Same story, same arc, same illustrations. Different prose, different safety steps, different teacher.

The tooling

Same stack as v1. The only voice swap: Matilda → Bill for Cypher (mature mentor register for the older tiger).

Read it / hear it

Per-chapter mp3s live inside the novel page. Audio files are also in /books/april-and-the-dragon-staff-v2/audio/. The master audiobook.mp3 has ID3v2.3 chapter markers + embedded cover artwork — Apple Podcasts, Overcast, and Pocket Casts will show a clickable chapter list.

Want to see the original? Click v1 in the tabs above.

The numbers

A four-sentence brief became a 59-minute illustrated audiobook in strict rhymed verse. Here's the count of everything that got generated along the way.

"The dragon staff has three bright wicks,
and fire is not for careless tricks."
"We hear you, April, loud and true —
now teach us what the dragon can do!"

…

"Now shake off the extra before the fire can grow.
Three flicks from the wrist and the excess falls free —
the burn-off comes FIRST. Mango — answer me."

                              (silence — two blank lines)

The night holds its breath. Not a bell. Not a sound.
Mango's wings are already mid-sweep off the ground.

The cast — and their voices

April

Sarah

Rainbow koala teacher

Mango

Jessica

Fastest fruit bat

Fern

Lily

Careful tree frog

Riddle

Laura

Questioning parrot

Dusty

Roger

Steady wombat

Laurel

Matilda

April's old teacher (Ch. 9)

Plus George (Narrator) — warm British storyteller, the constant voice across all 10 chapters.

Cost breakdown

Two ways to read this. Out-of-pocket is what actually hit a credit card. At-full-rates is what a fresh account with no credits or subscriptions would have paid for the same work.

What absorbed the cost. xAI's image-gen credit grant covers all 106 renders here (this book is xAI-only on images — Grok-Imagine nails the Pixar Storybook 3D aesthetic on the first try, so OpenAI didn't need to run). ElevenLabs lands at $0.015 per request on Sean's Creator subscription — the real out-of-pocket line on this book. 490 requests came from the call-and-response form: every voice switch is its own segment, so the segmentation count is denser than a prose audiobook. Claude Code is a flat monthly subscription that covers all the agent's planning and tool calls.

What a fresh account would pay. ~$35 at-full-rates is what you'd spend if you opened brand-new accounts at every provider and rebuilt this exact book. About a third less than Can't Have Nice Things (~$42) — chiefly because this book renders illustrations only with Grok-Imagine ($0.07/image) instead of OpenAI's gpt-image-1 ($0.04–$0.17/image), and because the audiobook is half the length.

Timing

Wall-clock time for each stage. Everything that could parallelize did. Total active runtime from "go" to a finished audiobook: roughly 20 minutes (plus a brief pause mid-way through chapter 10 to top up ElevenLabs).

The process

Ten stages, each a separate slash command or script. The early ones generate a menu of options; the late ones commit to picks and bake the cake. Two stages were skippable for this book — the style was pre-locked by a reference image, and the preview wizard was auto-resolved because the judges agreed by a wide margin.

The tooling

Every step is a shell-out, not a library binding — the pipeline is a stack of bash scripts and Python helpers. Sibling of choir and scriptorium.

Read it / hear it

Per-chapter mp3s live inside the novel page — each chapter has its own <audio> player above the verse. The chapter audio files are also in /books/april-and-the-dragon-staff/audio/ if you want to drop them into a podcast app or pull them down individually. The master audiobook.mp3 has ID3v2.3 chapter markers + embedded cover artwork, so apps like Apple Podcasts, Overcast, and Pocket Casts will show a clickable chapter list.