Chapter 4. Drawing Out Every Keyword — The Posture of First Extraction

Think of the moment you first open an IP document. A rulebook, several short stories, and a bundle of planning documents lie before you, and you must draw the whole web of relations in this IP out of them. Where do you begin, and what do you grasp first?

The name “first extraction” looks simple, but the posture of this stage determines the texture of Librarying as a whole. When the first stage is blurred, all three following stages blur; when it is firm, the next three become much lighter.

One principle governs the entire stage: gather first, organize later. A scientist who gathers data while trying to confirm a hypothesis falls into confirmation bias. First extraction is no different. Once you start judging, while pulling out a keyword, whether it is a hub or a leaf or how it will be grouped in second integration, you miss good keywords themselves. At the first stage, then, focus only on drawing things out and leave the remaining judgments to later stages.

This was precisely the posture used by the editors of the Oxford English Dictionary in the 1860s. They distributed tens of thousands of books among thousands of volunteers and had them write each occurrence of a word, with one line of source information, on a card ten centimeters wide. Only after roughly five million cards had accumulated did organization and integration begin. Each card became one row; the mountain of cards piled up before organization began is the result of first extraction. Librarying’s first stage follows that posture exactly.

This is why first extraction sets its acceptance threshold at 60 percent. Duplicates are acceptable; it does not matter if the same character enters twice under different forms, or if a category is slightly off. Items whose category is unclear are placed provisionally in the nearest category, and the decision is postponed to the next stage.

The point is not to miss anything. A keyword missed once will not be discovered again in a later stage. Unneeded items among those that enter can be removed in second integration, but an item that never entered at the beginning never enters at all.

How Many Keywords Are on One Page?

Let us begin with a small experiment. Take out one page from a rulebook, short story, or planning document for an IP you operate. How many keywords are on that page?

Different people give different answers. Some see five; others see twenty. The difference comes from how far one is willing to treat something as a keyword.

Consider a page in a rulebook. Suppose it describes, in one paragraph, a scene from the modern hunter-fiction worldview “Gongdong”: “Kang Siwoo took out his rank card before Gate 13. The rank field on his card, as a non-attributed person, remained blank. Yoon Seogyeong, an examiner from the Hunters Association, held a concentration meter to him, but its needle pointed to blankness and declared measurement impossible. Yoon recalled the mass re-awakening accident eighteen years earlier. While crack-light leaked out from beyond the gate, the blue light of a system window never appeared in Siwoo’s sight.”

Read normally, it seems like a natural descriptive paragraph. Seen through first extraction, however, it contains at least twelve keywords: two character names (Kang Siwoo and Yoon Seogyeong), one title (examiner), two places (Gate 13 and the Hunters Association), two objects (rank card and concentration meter), one time point (eighteen years ago), one event (the mass re-awakening accident), two proper terms (non-attributed and measurement impossible), and one operating principle (system window).

There may be more. If crack-light is a visual image repeated at every gate in that IP, it is description embedded in place but also belongs to mise-en-scène. If the blank field on the rank card repeatedly symbolizes non-attribution, it is an object and also mise-en-scène. The number of keywords that can be drawn from one page varies with the depth of the posture used to draw them out. Librarying’s first extraction seeks the deepest extraction possible.

Let us place these twelve keywords in the nine categories. “Non-attributed” and “measurement impossible” are proper terms requiring explanation for someone unfamiliar with the IP, so they are technical terms. Kang Siwoo and Yoon Seogyeong function as agents, so they are characters; “examiner” is also a character because it refers to an actual person in the text. If it were used as an abstraction—the Association’s system of ranks—it would be a concept. Gate 13 and the Hunters Association are background spaces, so they are places. The rank card and concentration meter are graspable things, so they are objects. “The mass re-awakening accident eighteen years ago” is something that happened at a particular time, so it is an event. The system window is an operating principle that displays abilities and ranks, so it is a mechanism. Crack-light is a recurrent atmospheric element at gates, so it is mise-en-scène.

Three distribution points deserve attention. First, “examiner” is categorized as a character because it refers to an actual person in this text; if it meant the abstract system of Association posts, it would be a concept. Second, the rank card is an object, but if its blank field recurs as a symbol of non-attribution, register it again as mise-en-scène in first extraction and decide in second integration. Third, where classification is unclear, the 60 percent threshold means placing it in the nearest category for now and moving on. This is the first exercise in turning one paragraph’s keywords into a cluster of rows in a five-column CSV.

Nine Categories: A Classification by Narrative Function

Librarying draws keywords into nine categories, called the nine categories in this book: technical terms, characters, places, objects, events, mechanisms, mise-en-scène, concepts, and values. Every keyword in an IP is placed in one of these nine, or sometimes two.

All nine divide according to “what function does this keyword serve in the narrative?” This is called classification by narrative function. Formal classifications such as alphabetical order, importance, or frequency cannot explain why those keywords belong together. Narrative-function classification can: all keywords in a group serve the same role in a story. Every keyword in characters functions as an agent, and every keyword in places functions as a setting. This shared function helps both when writing headword narratives and when output work uses the Vault.

The worldview axis introduced in Chapter 6 is the second classification criterion. The two criteria independently give each keyword coordinates: if the nine categories are the X-axis of narrative function, the worldview axis is the Y-axis of IP design. For now, remember only that the nine categories are a classification by narrative function.

The nine categories may be unfolded as three groups.

Proper vocabulary and people. A technical term is a proper word that exists only in the IP. Every term such as “black void,” “coffin,” “divine weapon,” “implant,” or “dimensional movement” that needs explanation for someone unfamiliar with the IP belongs here. It must be filled to saturation first, and it often forms the core of hubs. A character is every person-like being appearing in the IP: protagonist, supporting character, minor role, person in recollection, and even an unnamed title. This includes not only people but AI, spirits, and gods if they function as personhood. If the same person appears under several forms—Kang Siwoo and Siwoo in the body text—extract every form separately and group them only in second integration. A group of people such as a family, faction, or religious order is a character when it functions as a single personhood, a concept when it functions as an abstract system. If it spans both, register two rows.

Space and objects. A place is every space, from city, region, and dimension to building or one room. Draw out every place, whether directly shown or only mentioned. When hierarchy exists—from country to city to residence to a restaurant within it—draw it out flatly in first extraction and arrange the hierarchy in second integration. An object is every graspable thing: tools, equipment, items, food, materials, furniture, ornaments, and markers. Do not omit an object because it seems trivial; it may later become an important symbol. A mechanism is an IP’s operating principle or narrative device: game mechanics, magic rules, system functions, and principles of narrative operation such as foreshadowing, recurring motifs, or a narrator’s intervention. It is a category with many hub headwords. When object and mechanism are confusing, as with a magic wand, classify it as a mechanism when operating principle is central, and as an object when shape, material, or ownership history is central. Register both if both apply.

Time, atmosphere, and abstraction. An event is something that occurred in the work: episodes, flashbacks, and occurrences with a time point, outcome, and related people. Event classification organizes the IP’s time axis. If events are drawn out without omission in first extraction, second integration can arrange an event master and make the IP’s timeline visible as one line.

Mise-en-scène consists of elements that make an IP’s atmosphere: visual and auditory moods, recurring background images, symbolically recurring colors or sounds. If rain appears at every turning point for a character, “rain” is a mise-en-scène keyword. Because it is dissolved into atmosphere rather than appearing as a physical object in the body, it is often missed in first extraction. Give an AI work partner ample examples of mise-en-scène with the nine-category guide to improve accuracy.

Concept is an abstraction: the world’s setting principles, regulations, laws, social systems, institutions, and operating principles presumed by the text. The line “her dowry was five thousand pounds” presupposes the concept of marriage economics; “he was not qualified to receive a noble title” presupposes a system of noble titles. AI work partners often miss implied keywords, so people inspect them once more in macro review.

Value is what the IP conveys: thematic consciousness, emotion, relationships, recurring motifs, and abstract layers that form its core messages, such as revenge, sacrifice, freedom, and family. If a concept is “how the world operates,” a value is “what matters and what is right within that world.”

The consistent rule for blurred category boundaries is to place an item in both categories during first extraction and decide in second integration. The first-stage threshold is 60 percent, and omission is the greater loss.

One Row in a Five-Column CSV

Keywords drawn out through the nine categories are arranged one per row in a five-column CSV. A row holds one keyword, with information distributed across five cells.

IDX is the serial number that uniquely distinguishes a row; category shows which of the nine categories it belongs to. Keyword records the form exactly as it appears in the source, whether Korean or English. Description is a one- or two-line definition. Source records the document and position in which it occurs.

The result of first extraction for one IP is a table containing 5,000 to 15,000 of these five-column rows.

Because the keyword field preserves source forms, “black void” and “black smoke” appear in separate rows if both occur; deciding whether they mean the same thing waits for second integration. Translation candidates, such as English names and romanization, are decided not here but in the second stage’s thirteen columns. The description is a brief definition; where unclear, write only “presumed to be ○○” and refine it later. The source field, by contrast, must be filled most accurately at first extraction: if its document, chapter, and line position blur, the keyword cannot be found again in the source at the next stage.

What People Must Draw Out Themselves

AI work partners handle the core of first extraction, but several strands still require people. The first is fine classification judgment: deciding cases that confuse object and mechanism, or event and concept. With the nine-category guide, AI generally classifies well, but people know which side fits this particular IP when the border is unclear. If, at the start, you decide standards such as “magic-wand types are mechanisms in this IP” or “title names are characters,” write them in the guide and deliver them with the work; first extraction then retains a consistent texture.

People also identify truly new domains. When a newly appearing keyword fits none of the nine exactly, people decide whether a new category is needed. In practice, however, it is first placed in the closest category, and the question of a new category is settled during second integration.

People next track keywords appearing only as pronouns or demonstratives: the “he” in “he entered,” the “there” in “they met there.” Infer their category from context and extract them with the uncertainty marked in the keyword field—“he (presumed: character),” “there (presumed: place)”—then verify them in second integration.

A character or place appearing only as a pronoun in one work often receives a name only in another: the “he” of the first work may become Kang Siwoo in the fifth. If first extraction misses that “he,” Kang Siwoo in the fifth work remains forever unconnected to the first. AI work partners often miss this, so after extraction a person scans the same document once more and marks pronoun and demonstrative keywords. This takes about fifteen to thirty minutes per document.

Finally, people perform macro review. After first extraction, scan the result table to see whether one of the nine categories is abnormally empty—perhaps AI missed the category altogether—or whether one category is swollen, signaling a need for further distinction within it. This is not a close reading of the whole table; it catches outliers by looking at category totals and takes about thirty minutes.

Delegating to AI Work Partners

This is where AI work partners perform the main work of first extraction. In the past, people divided a large manuscript into a manageable number of lines and had it read piece by piece. The current standard differs. Input text is first normalized as a whole and prepared in a form that preserves which file and line range each keyword comes from. The AI work partner does not stop after reading only the beginning out of concern for length limits; it concentrates on exhaustive extraction from the full text provided.

A chapter of a rulebook, one short story, or one planning document is one unit. Give that unit the nine-category guide and five-column CSV format, and the AI work partner reads it, extracts keywords by category, and returns a five-column CSV. A short story’s body, work description, glossary, and author note are not handled identically: body text is treated through usage; the description as a high-density source of settings; the glossary as term-definition pairs; and meta-commentary is generally excluded. In rulebooks, rules definitions become technical terms and mechanisms; numeric tables yield not just numbers but units and context. This is why the type of input must be stated first.

Delegation proceeds in parallel. If an IP has fifty input documents, multiple AI work partners process several documents at once. Five partners processing five documents finish five documents in the same thirty minutes in which one person processes one. Fifty documents, grouped five at a time, finish in ten thirty-minute rounds. The person works on other things meanwhile. This parallelism is a major reason Librarying finishes in nine to sixteen hours: what takes a month to read manually across fifty documents falls to about two hours when delegated in parallel. Speed alone is not enough, however. After extraction, check that sources do not cluster only at the beginning of input documents. If there are scarcely any sources from later line ranges, extraction probably stopped midway.

The four axes of this collaboration are these. Input consists of one document unit, the nine-category guide—definitions and IP-specific fine decisions—the five-column CSV format, and the first-stage acceptance posture. That posture permits duplicates and typos and does not cling to polish. Output is one five-column CSV table; intent is time reduction. Work that takes a person an hour to read and draw out ends in roughly five to ten minutes, and running fifty units in parallel greatly reduces actual human effort. The expected result is 5,000 to 15,000 rows in about thirty minutes to two hours.

The limits are areas delegation cannot reach: implied keywords, often concepts and values not explicitly stated in the body; characters, objects, and places appearing only as pronouns; dates and chronological references such as which year “that spring” means (extract the form as written at first stage and decide later); and IP-specific jokes or citations, such as one work citing another or one line mirroring another character’s line. These are materials that exist only in the mind of someone who has read five years of work. They are the four areas people must pass through once more in first extraction. Their combined time is about one to two hours per IP, and the full first extraction still ends in roughly one or two hours.

Moving Text into a Knowledge Structure

To creative practice, first extraction feels like pulling out keywords. Viewed more broadly, it moves text into a knowledge structure. The problem is close to one long handled by systems such as TEI in digital humanities: names, place names, dates, annotations, quotations, and edition information mixed into literary text are not left buried in sentences, but marked in a structure that can be found and compared later.

Librarying does not make XML markup as TEI does. At a much lower level, it reads work bodies, setting collections, author notes, and glossaries, and converts them into candidate headwords such as character, place, event, mechanism, and concept. The body gives usage; a setting collection gives definition; an author note gives intention; a glossary gives existing naming decisions. Different input types yield different kinds of knowledge.

This perspective clarifies the attitude of first extraction. We are not summarizing the source. We are drawing out knowledge units scattered within it and turning them into material that can be connected later. That is why sources and line ranges are preserved carefully, and why bodies, descriptions, glossaries, and author notes are not flattened into one method. First extraction’s output is not a simple list; it is the first conversion table from text into knowledge structure.

The Output of First Extraction

When the first stage ends, you hold one five-column CSV for the IP: 5,000 to 15,000 rows, ordered by the nine categories, with one keyword per row. At ten thousand rows, no one can read the table all at once. Fortunately, the first stage does not require that: macro review only needs category totals, and close inspection begins with second integration.

After first extraction, choose one of the nine categories and scan about its first fifty rows. You will feel the posture with which the IP’s first extraction proceeded: which categories the AI work partner drew well and which awkwardly. Reading fifty rows takes about fifteen minutes, and that time lightens the decisions of second integration.

A common concern follows: what if the AI work partner extracts keywords incorrectly, and an inaccurate five-column table contains 5,000 rows? The answer is simple. Sixty percent accuracy is sufficient at the first stage, so some misclassification within that sixty percent is acceptable because people decide once more in second integration. In practice, AI work partners exceed 80 percent in first extraction and approach 90 percent when the nine-category guide is well refined. Since a stage needing 60 percent receives 80 to 90 percent, first extraction passes consistently. This generous threshold belongs only to the first stage. The next stage, second integration, requires the much stricter 95 percent. Acceptance thresholds differ by stage so that first-stage generosity does not flow forward unchanged.

Closing First Extraction

Gather first, organize later. Perfect organization is not the goal. The goal is that nothing be missed; later stages do the organizing.

Chapter 5 is second integration: the process in which the first stage’s 5,000 to 15,000 rows are grouped into 1,000 to 2,000 headwords. Its 95 percent threshold is demanding, and human decisions become central.