← All work
Generative AI Agents · Prototype · 2026

Scriptorium

Agents plan, write and illustrate novels, novellas and short stories on a Vonnegut fortune curve, with a deterministic character graph that keeps a cast consistent across a thousand scenes.

Scriptorium
Year
2026
Status
Prototype
Category
Generative AI Agents
Role
Architect & Lead

Key metrics

15
Agent skills
1,000+ scenes
Scale
90–150k words
Novel budget

Case study

Scriptorium

A loop of agents that writes a whole book on a fortune-shape spine (Vonnegut / Reagan), illustrates it, and lets you watch the agents work. It runs as a closed beta at script.macleodlabs.com.

The problem

A language model can write a luminous scene. It cannot hold the evolving social world of a novel in its head across a thousand scenes: who cooled toward whom, who quietly betrayed whom four chapters ago, which faction is slowly turning. Running recaps work at fourteen scenes and fail at novel length.

What it does

  • The shelf. A library of every story, from short stories to novels, each with a generated cover, genre tags and progress.
  • The editor. A wizard that takes a premise from "surprise me" to an approved story plan.
  • The reader. Paper-grain, swipe-to-turn pages with the protagonist's fortune curve in the margin, plus PDF and Kindle EPUB export.
  • Arc and map. Every character's journey charted from good to ill, planned pacing against the tension each scene actually delivered, storylines showing who is on stage where, and a hand-drawn world map of every location in the book.
  • The chalkboard. A live generation board: character fortune curves drawn in chalk as each scene is sealed, cast state, a location map, storyboard panels and an inspector for every gate decision.

How it works

  • Grown, not sliced. One recursive refine step turns a storyline into chapters that fit the arc, scenes that fit each chapter, and beats. Counts are emergent, bounded by a soft genre word budget (90–150k words for science fiction and fantasy).
  • Story shape. A CP-SAT arc-fitter places each character's independent fortune curve so their lives intersect at the right beats, solved per chapter at novel scale.
  • A character and world graph. A sparse, event-sourced model of valence, trust, power, beliefs, moods and factions. The model writes the prose and proposes what changed; the graph remembers everything, ripples consequences to characters off-stage, and hands the next scene exactly the slice it needs. Read the design doc.
  • Agents. Fifteen single-purpose skills run as Claude Code subagents on a subscription, in a generate, audit and repair loop that writes prose, per-scene panels and a cover.
  • Resumable. Novels write chapter by chapter with a checkpoint after each, so an 80–120k-word book never runs in one pass.

Tech stack

PythonFastAPIWebSocketsClaude Code subagentsOR-Tools CP-SATNano Banana

Gallery

Other 2026 work