Quill

Quill is an exploration of how to coordinate AI content generation across a connected work. A picture book provides a concrete test: its story needs to hold together, recurring characters and objects need to remain recognizable, and each illustration needs to leave a usable place for the words.

Fisher’s interest is the pipeline that connects those decisions. A plausible page produced in isolation is insufficient evidence that the next page will agree with it. Quill makes narrative plans, visual definitions, reference images and observations of finished art into inputs that later stages can reuse.

One book, connected artifacts

The worked example is The Blueprint and the Breeze, an original story about a kite trapped in an oak tree. Its arc turns repeated retrieval attempts into acceptance that the tree can keep flying the kite. The kite, notebook, toolbox and park recur across the pages; they are useful tests of how identity and changing circumstances travel through generation.

The example preserves planning calls recorded on August 13, 2026, and the resulting reference sheets and page illustrations. Its image-generation and seed-selection calls were not recorded, so these artifacts do not establish which earlier images reached any particular image call. The pipeline map describes the current implementation; the book supplies a concrete set of artifacts to follow through it.

A story before a stack of pages

The initial brief supplies a premise, reading duration and three roles: author voice, illustration style and art direction. An arc establishes the protagonist’s want, obstacle, turn and resolution. Cast definitions establish who appears and how each character should be described visually.

The author then writes the story as one continuous piece. Length correction cuts or develops whole passages before the story is divided into pages. Pagination chooses breaks and scene groupings; art direction adds subjects, shots and lighting to those pages.

This order addresses a functional problem with composing prose into preassigned page slots: locally reasonable beats can repeat, and transitions can disappear between independently written pieces. The story becomes the material to arrange, rather than a collection of empty boxes to fill. Pagination is intended to preserve its wording. Severe under-generation can trigger a retry or rejection; a word-count warning alone cannot prove that every word survived unchanged.

Identify what should persist

The world stage extracts important props and recurring places from the written story. Each gets a dense visual definition: material, color, distinguishing marks or the features that make a room recognizable. A grounding pass folds explicit visual facts from the story back into the definitions used for composition and reference generation.

Page composition then describes the painting and identifies the entities needed on that page. It also plans a quiet region for text and approximate subject positions. The recurring-reference pass determines which settings and objects justify reusable visual references.

For the kite, the world record specifies ruby-red diamond-shaped fabric, birch spars, a white ribbon tail and a plastic spool. The recorded composition input includes that description, and the page-two output repeats its construction while placing it between oak branches. This is a visible producer-to-consumer connection, not a field inferred to be useful because it exists in a schema.

The result is a set of relationships between the book’s narrative, its entities and its pages. A prop definition supplies identity; a page’s composition supplies the moment. They serve different purposes and need to reach the right downstream call.

References are inputs, not a global mood board

After draft approval, Quill generates a style image, character turnarounds and references for recurring settings and qualifying props. Those assets are selected for the pages that actually need them. A major prop can earn a detailed sheet; a lesser recurring prop can get a simpler one.

Reference scope matters. An image supplied as context can become an instruction to include its contents. Sending every object to every page may paint a recognizable object into a scene where it does not belong. The pipeline therefore carries page-to-reference relationships instead of attaching the entire asset collection indiscriminately.

The kite’s generated sheet shows its shape, spars and tail without committing to a particular tree scene. The character turnaround and illustration style provide other reusable inputs. Later page art shows the kite lodged in the oak and the character returning to that setting; those outputs can be compared with the definitions and sheets without assuming that every planned detail survived.

The recorded page-two composition explicitly lists its prop dependency:

{
    "page": 2,
    "props": ["Toby's Crimson Kite"],
    "quiet_rect": {"X": 0.08, "Y": 0.68, "W": 0.84, "H": 0.28}
}

These fields are a selected excerpt of the actual composition artifact. props names the asset relationship; quiet_rect describes intended space in normalized page coordinates. Neither is a claim about what the image model ultimately painted.

These baseline assets establish identity. They do not by themselves capture everything that emerges during illustration: the scale of an object, its staging, or what it looks like after the story changes it.

Carry forward what was actually painted

Pages are illustrated in sequence. Each completed page joins a pool containing its scene, its actual image and a short account of painted facts produced by a vision read. Before the next page is drawn, a selector can choose up to four relevant earlier pages from that pool.

The selected images and observed facts accompany the new page’s baseline references and composition. The relevant predecessor can be a page from the same scene or an earlier place the story returns to. It is a bounded selection from already rendered evidence, rather than a fixed rule that always reuses the first page.

This addresses continuity that descriptions and turnarounds alone can miss. A recurring object can keep the same name while being illustrated at a different scale or with a different construction on each page. Reading and reusing actual images gives later calls more concrete evidence of what needs to persist.

That mechanism can also carry a changed appearance forward, but it is not a formal object-state system. Temporal order lives in the story and page sequence. There is no explicit per-object event ledger or versioned chain of state assets, and a redraw does not automatically rebuild dependent later pages. Continuity remains something to inspect, not a result guaranteed by supplying references.

Place words against the result, not the plan

The composition’s quiet rectangle and subject positions describe intent. They cannot establish where the illustrator actually put a face or whether the intended text area became busy.

After illustration, Quill measures variation in the image and asks a vision model for keep-out regions and usable text areas. The same vision read also supplies continuity observations for later pages. A separate reconciliation pass can adjust words that plainly contradict the finished picture, while attempting to preserve the voice and meaning.

The reader chooses a book-wide type size and starts with a fast placement. A background search evaluates alternatives against contrast, occlusion, composition and typography, then saves the result. Human placements take precedence over computed ones.

The important dependency is on finished pixels. Typography has to work with the art that exists, even when it differs from the requested painting. An automatically accepted placement still needs visual inspection; an internal score is not proof that a page reads well.

An inspectable generation pipeline

The worked example also exposes a limitation: its recorded cast calls the lead Calvin, while the story and world use Toby. The discrepancy is retained in the evidence. An image can look consistent while the structured artifacts disagree about the identity it represents.

Quill’s tracing mechanism records prompts, responses and labeled reference-image metadata when a call is trace-bound. Those records make it possible to inspect what an earlier stage supplied and whether it actually reached a later call. A field present in a schema but unused downstream is not a functioning connection.

Coverage needs inspection too. This book’s planning trace does not include its image calls. A record of a world definition reaching composition proves that connection; it does not prove the later model received the intended reference set. Complete instrumentation is part of making the pipeline explainable.

The pipeline is ordered so narrative planning and draft review precede the book’s image-generation work. Sequential illustration improves access to earlier evidence at the cost of parallel throughput. Failed reference selection or vision reads can degrade continuity without necessarily stopping the book, which makes the recorded calls and final-page inspection part of the engineering work.

The picture book is the worked example. The broader investigation is how to make a multi-stage generative system carry identity, context and observations forward without treating each output as an unrelated completion.