Chapter 12. Working with an AI Collaborator — A Book Made by Two Sets of Hands

Picture a film set. It is not a one-person operation with only one camera, one microphone, and one lighting kit. A cinematographer, lighting director, sound director, costume, art, and makeup teams work together with a single director at the center of every decision to construct a scene. The director may not hold the camera or move the lights, but the substance of every decision comes from the director. The director owns the substance of decisions; the other hands carry out their execution. This division is why a LOREBOOK project can be completed in nine to fourteen months. It is not because people suddenly write faster.

It is because the work performed by people and the work entrusted to an AI collaborator are divided into different kinds of work. Mapping 1,500 entries to book parts takes a person a week by hand, but automatic mapping takes one or two hours. Converting frontmatter by hand takes another week; automatic conversion takes five minutes. When differences like these accumulate in five or so places, almost a year disappears.

Not every task divides this way, however. Work that requires human judgment should not be delegated to automation; when it is, the result becomes awkward. This chapter sets out the structure of the division: which work moves into the automatic domain, which remains human work, and which collaborative tools enter where the two domains meet.

The automatic domain — conversion, candidates, and checking

Tasks that can be delegated well to an AI collaborator share three characteristics: they are mechanical, they require large sets of candidates, or they can be checked automatically.

Mechanical conversion is the lightest domain. Converting the thirteen Librarying columns into LOREBOOK frontmatter is a representative case, alongside automatic directory creation in Phase 1, automatic generation of indexes and glossaries in Phase 5, and the web-conversion build in Phase 6. Work that takes a week by hand is completed in five to ten minutes, and human attention is needed only for review—about thirty minutes for a sample of fifty items.

The most substantial domain is producing large sets of candidates. Candidates for part mapping, synonym clusters, categories and labels, illustration specification cards, short-story quotations, and first drafts of multilingual translations all belong here. Given material for 1,500 entries and guidelines, an AI collaborator can make a candidate matrix in one to three hours; a person then spends four to ten hours reviewing it and making decisions. Because this category accounts for roughly 40–50 percent of LOREBOOK work, the collaboration between generating candidates and deciding among them is the largest part of the process.

Automatic checking operates in consistency verification and tone-consistency review. It catches clear violations: dead links, duplicate entries, frontmatter mismatches, five absolute prohibitions, and noun searches. A violation report appears in roughly one or two hours; people decide, entry by entry, whether to revise, rewrite, or work around each issue. Tools that check whether a part's tone mapping has been properly applied belong to the same category: they identify paragraphs that stray from the assigned tone. Usually thirty to fifty cases are identified, after which a person decides whether each needs partial revision or a complete rewrite.

Together, these three categories account for roughly 60–70 percent of the project time, and all share one quality: they are work of organizing material.

The human domain — decisions that make the book what it is

All work decided by people gathers around the parts that determine the book's identity. If automation handles the execution of a decision, people determine its substance.

People choose the book's structure—whether it will have ten, twelve, or fourteen parts—and set the balance of page counts and the priorities of those parts. Decisions to merge or divide also belong to people. Whether several entries with the same coordinates should be integrated into one entry or separated, and whether a mega-hub entry should be divided across several locations, determines the shape of the book. The threshold for accepting a tone is the same: people judge whether a part's prose fits its tone mapping and thereby establish the book's voice. On top of this come the timing of illustration commissions and review of the integrated build, which determine the moment at which the book reaches readers as a finished object.

All of these decisions require the judgment of someone who knows the IP deeply. An AI collaborator can offer candidates, but the person who knows the IP's future operating plan determines its part structure; the person who knows its readers' patterns of use decides on integration and division; and the person who knows its tone judges whether the voice is acceptable. These five categories may make up only 30–40 percent of project time, but in terms of decisions that govern the book's identity, almost all of them remain in human hands.

The four-axis collaboration tool

The tool for organizing how the two sets of hands meet at every stage is the four axes of collaboration: specification, intent, expected effect, and limits.

Specification is the format of input and output: what material will be supplied and what form the result should take. Intent is the reason for carrying out the task collaboratively, and includes aims such as saving time, consistency, handling 1,500 units, and securing a first draft. Expected effect combines the time saved by collaboration with the time required from people. Limits identify the parts delegation cannot reach—the domains that people must decide by hand.

When these four axes are arranged in the same form for every Phase, readers can apply collaboration to their own IP with their own tools by laying out the same information each time. Consider how the four axes work in the example of Part 4, Geography, for the hypothetical IP “○○.”

Specification defines the input and output formats. The input consists of material for ninety entries mapped to Part 4 Geography in the Vault; the part's tone mapping (field-report style as the primary voice, catalog style as the secondary voice); the prose guides in Chapters 7–9; the Part 4 length guide (fifty pages, twenty to thirty entries); and five map illustration-specification cards. The output consists of first prose drafts for the ninety entries in Part 4, the README files for Part 4 and Chapter 5, and material for reviewing the illustration specification cards.

Its intent has three parts: saving time, consistency, and handling 1,500 units. Writing ninety Part 4 entries by hand takes roughly six weeks, whereas an AI collaborator can produce first drafts in about five days. The tones of the ninety entries remain consistently aligned with Part 4's tone mapping—primarily field-report style—and all ninety entries are organized in the same way as the other eleven parts.

The expected effect totals about five weeks: approximately five days for the AI collaborator's first draft, three weeks for human review and refinement, and one week for a Part 4 consistency check.

There are five limits that remain human responsibilities: deciding Part 4's priorities, deciding how to merge or divide the six style zones, determining descriptions in a field-report voice suited to the IP's cities, deciding when to commission illustrations, and making the GO/NO-GO decision in the Part 4 master review.

This four-axis matrix organizes collaboration within a single part. If readers establish the same four axes in advance for a part of their own IP, the work becomes cleanly structured; when trouble arises, they can inspect which of the four axes has slipped out of alignment. The integrated four-axis matrix for all seven Phases is collected in Appendix C.

Five stages of review — in chronological order

Although collaboration enters every Phase, there is a separate integrated review sequence that ties the entire project together. In chronological order, its five stages progressively shift their emphasis from automation to people.

Automatic checking comes first. Review scripts initially catch dead links, duplicate entries, frontmatter consistency, five absolute prohibitions, and noun searches. It finishes in about one or two hours; doing the same work by hand would take a week. Next comes human random-sample review. Thirty to fifty of the 1,500 entries are opened at random to repair subtle issues that automatic checks cannot detect. This takes about one to two days.

Third is hub-priority review. Since roughly 100–150 hub entries among the 1,500 are the LOREBOOK's core, they receive the closest review, taking about one to two weeks. Above that sits a series-wide tone-consistency review lasting roughly three to five days, which examines whether tonal shifts between parts feel natural. Finally comes integrated-build review: a pilot build—web or print PDF—is read through as a single book and completed over one to two weeks.

All five stages together take roughly one to two months. Automatic checking excels at catching clear violations quickly, while human review excels at examining subtle mismatches that automation misses and the impression received while reading the book continuously.

Dividing work among multiple AI collaborators

A LOREBOOK project can use one AI collaborator or several at once. The difference lies in the balance between working time and consistency.

The first arrangement is parallel work by part. In a twelve-part structure, eleven parts other than the pilot part are divided among eleven AI collaborators, each producing first prose drafts for one part. One AI collaborator working through all eleven parts sequentially takes about one to two months; eleven working at once collect their first drafts in about one to two weeks.

One major problem follows: consistency between parts. Even when eleven AI collaborators receive the same guide, their tones, SSOT quotation patterns, and avoidance of absolute prohibitions differ subtly, increasing the cost of consistency review. A balance is therefore recommended. Have a single AI collaborator build the pilot part to establish the tone guide; then, in Phase 3, divide the eleven parts among three or four AI collaborators. Each takes three or four parts, and assigning parts with the same mapped voice to the same collaborator naturally preserves consistency.

Parallel work by chapter within a part is also possible. If a part contains five chapters, five AI collaborators can each draft a chapter at the same time. This reduces a part's work from about a week to two or three days, but consistency between chapters must be designed in the same way. The recommendation is to have a single AI collaborator produce the first chapter—the anchor chapter—and pass it to the remaining four chapters as their guide.

Where time savings occur

The areas in which a LOREBOOK project is shortened from eighteen to twenty-four months to nine to fourteen months are clear. Because LOREBOOK directly receives Librarying's thirteen-column frontmatter, the time for writing frontmatter falls almost to zero (Chapter 5). The two coordinates—nine classifications and worldbuilding axes—automatically produce part mappings, reducing a week to one or two hours (Chapter 6). Once first prose drafts for the eleven parts are made in parallel by part, working time falls to one-third or one-quarter. Appendices such as glossaries and indexes are generated automatically from frontmatter, reducing a month to a week. And the first stage of the five review stages runs automatically, shortening the time needed for human review.

Almost a year disappears across these five areas, but it is important that human work has not disappeared. Work of organizing material has moved into automation, while work of deciding remains human work. Automation does not replace people; it changes the domain in which people work.

The reader's first task

After closing this chapter, the reader's first task has four steps.

First, check Librarying. Because the input for this book is a Vault after Librarying is complete, enter LOREBOOK work directly if it is finished; if it is not, begin with the material-organizing stage. Second, write the Bible agreement document. Organize the eight decisions discussed in Chapter 10 in one document, about ten to twenty pages long. Third, choose an AI collaborator. This chapter does not name a specific model, so readers should choose a tool they know well in their own environment. Finally, build one pilot part. Select one of the twelve parts as the pilot, begin work, and follow the procedure in Chapter 10. When the pilot part receives a GO decision, enter Phase 3.

These four steps are the reader's first task. Once the first step is complete, one foot is already in the work. Nine to fourteen months later, a book will arrive on the reader's shelf.

Copyright in this work belongs to Kim Dong-eun WhtDrgon and MEJE Works Corp. Without the copyright holders' prior written consent, no part of this work may be reproduced, distributed, transmitted, displayed, performed, broadcast, translated, adapted, or otherwise used.