Chapter 9. Fourth-Stage Narrative Writing: Breathing Life into the Vault

Look up the headword “Rain-Prayer Rite” in the Eastern-dynasty world Usabu, and you find the definition “a state rite praying for rain during drought.” The detailed description organizes its procedure, conditions, and consequences. But can a writer compose the scene with only that? To write the noon when someone first prayed toward a dry sky with cracked fields at their back—the feeling somewhere between bewilderment and desperation—definition and detailed description alone do not give the hand a place to begin.

Fourth-stage narrative writing is the stage that writes that narrative.

Why narrative is needed: why definitions are not enough

Recall the definition made in Chapter 5: “A rain-prayer rite is a state rite praying for rain during drought. It becomes a stage that tests the sovereign’s virtue and legitimacy.” It is accurate, self-contained, and immediately intelligible when cited in another headword’s detailed description: a good definition.

Yet it is not enough. A writer trying to portray a person conducting a rain-prayer rite for the first time cannot immediately write a scene after reading it. What atmosphere the rite has, what it feels like, how people in this IP receive it, and how the person who first performed it changes afterward all remain outside the definition.

What is needed here is a scene-making point. A scene-making point is not a finished manuscript that directly turns a headword into novel prose or a storyboard. It is a creative landing point that definition and detailed description alone cannot grasp. It is smaller than an event and larger than a shot or cut: not the entire Rain-Prayer Rite event, but the noon moment of first looking up at the dry sky with cracked fields behind one’s back; the moment a king prays for rain while the crowd already foretells failure in silence; the instant no wind rises even after the officiant’s invocation ends.

This unit is translated differently by medium. In web novels it becomes a scene or point-of-view scene; in webtoons, a scene made from several grouped cuts; in film and drama, a scene or sequence. In games, it becomes an event scene, cutscene, quest scene, or sometimes a key stage moment. The name changes, but the function does not: it brings a headword’s concept down into a moment readers, viewers, or players can actually experience.

Scene-making points have several types. First is sensory: the smell of earth in cracked fields, white noon light, a dust-covered altar, a crowd making no sound—these establish the headword’s sensory temperature. Second is actional: an officiant raises an empty bowl, a sovereign kneels on dry ground, subjects look at a water gate rather than the sky—moments when the headword appears through bodily action. Third is relational: who performs the rite before whom, who believes or doubts it, and how the positions of their gazes divide.

Fourth is conflictual. When the rite is not simply prayer for rain but a stage testing the sovereign’s virtue and legitimacy, the headword becomes pressure rather than explanation. Fifth is transformational: after the rite ends and nothing happens, a person ceases to trust the sky; or a single late thread of rain hardens the order of power again. Not every type must be filled each time, but an important hub needs at least one scene-making point before a writer can turn it into an actual scene.

Encyclopedia-making has long distinguished definition from exposition. A definition describes category and characteristics logically; exposition tells how its subject actually works, the context in which it sits, and what makes it interesting. A good encyclopedia entry begins with definition and continues into exposition. In Librarying, the %% LLM %% block holds exposition. Definitions and detailed descriptions are logical specification; the LLM block is narrative showing that a headword is a living presence inside the IP world.

When this narrative exists, two things change. First is the experience of an IP creator consulting the Vault. A Vault with definitions alone is a dictionary; a Vault with LLM blocks filled is an IP-world memory repository. Searching a dictionary and walking through a memory repository are wholly different experiences. Second is the quality with which an AI work partner uses the Vault as context. With only a definition of “Rain-Prayer Rite,” AI produces a general-knowledge answer; with an LLM-block narrative, it answers with this IP’s particular texture. The density of the Vault determines the density of the output.

I have seen most clearly, in games, that a collection of individual headword narratives can make a world come alive. Elden Ring and Dark Souls conserve their main-story cutscenes but embed fragments of world history in individual descriptions of weapons, armor, and talismans; players gather the pieces and reconstruct the world in their minds. I take this as a model for fourth-stage narrative writing. If each headword is built like such an item description, a writer may read the Vault in scattered order and still find that its pieces gather into a living world. Hollow Knight follows the same principle: its in-universe Hunter’s Journal, filled as enemies are defeated, records not only enemy information but the hunter’s gaze and the world’s history across 168 entries, making the catalog itself an entrance into the world.

The structure of fourth-stage writing

Fourth-stage writing proceeds in four steps.

First, write hub headwords directly. The producer writes five to ten representative entries among the 100–150 hubs. These narratives become the tone standard for all subsequent writing. A headword-heavy IP may need only five; an IP whose tone is hard to capture may require close to ten.

Next, delegate the remainder to AI work partners. Divide remaining hubs, middle-tier entries, and leaves into batches. Batch size and concurrency are not fixed values: a batch can contain more light entries, but should shrink when it includes high-weight entries such as protagonists or key mechanisms.

Five review stages follow. Do not inject delegated output into files as received; only narratives that pass all five stages enter the LLM block.

Finally, update the manifest. The manifest summarizes build results. Recording progress after every batch makes it possible, when work pauses and resumes, to know which batch should continue next.

Writing five to ten hubs directly: finding the voice

The five to ten hubs a person writes directly are not simply high-quality narratives. They are the process by which the producer finds this IP narrative’s voice for the first time.

The same content becomes completely different text depending on the writer’s point of view and style. A narrative describing a rain-prayer rite might have the voice of an annalist neutrally recording ritual procedure; the inward voice of a person facing, for the first time, the bewilderment and desperation of conducting it; or the voice of a storyteller handing down this world’s ancient belief in Heaven’s mandate. Choosing the voice is not a technical decision but a creative one, and it is made in direct hub writing. The first three entries test the voice; after five, direction emerges; by ten, this IP narrative’s voice is complete.

Direct hub writing therefore receives the greatest care: twenty to thirty minutes for one headword. Five produce a minimum tone standard; close to ten produce a reference casebook crossing the worldbuilding’s many axes. This time must not be spared. Strong direct hub writing yields two effects: the quality of delegated results improves because AI work partners receive hub narratives as examples from which to learn the pattern; and the fourth of the five review stages, tonal consistency, becomes far lighter because a clear standard makes departures easy to detect.

Four tiers of length for character headwords

Fourth-stage writing applies different lengths according to a headword’s weight. Character headwords in particular are divided into four tiers.

Protagonists—the work’s core point of view and axis of conflict—receive six to eight paragraphs of free narrative. Include narrative significance and symbolism, character relationships, established appearance and affiliation, and the role this person plays inside the IP world.

Major characters—those involved in many events and affecting the narrative—receive three to five paragraphs. Supporting characters, who appear in some events or scenes, receive one to two paragraphs. Minor characters, whose names occur only once or twice, receive one to two sentences giving only their first appearance and brief role.

Why give a minor character only two sentences? Minor characters are leaves, and spending hub-level effort on a leaf destroys balance. Differences in headword length inside a Vault show their actual weight in the IP. If all headwords receive equal space, ironically, nothing is emphasized.

Wikipedia’s distribution of biographical entries follows the same four tiers. If “Shakespeare” is a long narrative exceeding ten thousand words, Horatio, a major character in Hamlet, has an independent document of around a thousand words; Francisco, the guard who appears briefly in Act 1, Scene 1, has no independent document at all and receives two lines in a character-list article. Length must differ for weight to become visible through length, and the same four-tier logic operates in Librarying character headwords.

Apply these four tiers to the Usabu IP now being worked on. [[Yacheonja]], the exception who reads the sky with the naked eye and correctly predicts rain, is the protagonist. Since that accuracy is the single variable that becomes treason and the axis of narrative conflict, write him most deeply in six to eight paragraphs. The Usabu Grand Academician at the other end of the same rule has protagonist-level weight and is written accordingly. Major characters are Usabu officials directly involved in prediction, water, and ritual, and powerful families controlling reservoirs and waterways; write them in three to five paragraphs. Supporting characters are lower Usabu officials and local clerks in upper towns appearing only in some scenes; write one to two paragraphs. Minor characters are people of lower towns, named only once or twice, who trust the heavenly signs, sow seeds, and collapse in drought; record only first appearance and role in one or two sentences. The same character pool divides into four lengths according to weight.

This tier division applies together with Chapter 8’s hub/leaf division. A person who is both protagonist and hub receives the deepest, longest treatment; a person who is both minor and leaf receives the shortest. The tiers and hub/leaf can sometimes diverge—for example, a minor character may receive many citations because they connect to an IP’s key mechanism—and then hub judgment takes precedence.

Structuring headwords other than characters

Other categories also contain structured information within their narratives.

For a mechanism headword, first unfold how the mechanism is experienced inside the IP world, then add a “Constraints” section. List what the mechanism can and cannot do, with sources. Put speculation or interpretation in the free-narrative area, and put only rules explicitly stated by sources in Constraints. This matters because a mechanism’s constraints create narrative: only by knowing what it cannot do can one judge what is possible in a scene where it appears.

For place and organization headwords, first unfold the role of that place or organization inside the IP world, then add a “Components” section. It lists people belonging to the place or organization and related mechanisms.

An event headword contains its position in the IP timeline, the people involved, and the changes it brought to the IP world. Causality matters especially here: why the event occurred and what changed as a result are the section’s core.

These structured sections are necessary for later work to retrieve information accurately from the Vault. With free narrative alone and no structure, tools make errors when extracting information from prose.

Story digest: a core input for Stage 4

One of the key inputs given to an AI work partner in fourth-stage writing is the story digest.

What should accompany a delegated headword? At a minimum, its definition and detailed description. But with only those, AI writes from general knowledge rather than this IP’s context. Give it only the definition of “Rain-Prayer Rite,” and it writes about rain rituals in East Asia generally; it cannot know what the rite means in this IP’s world.

The story digest solves this. It is not a short summary of original documents; it is an interpreted-context file for Stage 4 writing. The producer rereads materials such as manuscripts, scripts, and rulebooks, then compresses scene-by-scene plots, character relationships and movement, a key-event timeline, and recurring symbols and motifs into one body. Its length generally aims for about one fifth of the original, but conserving tokens is not the point. The point is to interpret why an event matters, which sensations recur, and where relationships turn, making material for fourth-stage prose. Given alongside a headword, it lets AI write within this IP’s particular context: it knows how the Rain-Prayer Rite happened to a particular character here and how that experience was described in the narrative.

Its contents are clear. Scene plots should not say merely “A did B,” but “A did B, therefore C changed,” exposing causal links. Character relationships include both their initial and final states. State recurring sensory motifs—colors, smells, sounds—and symbolic patterns. Order the key-event timeline by causality, not chronology. It is equally clear what not to include: do not paste original sentences verbatim. Excerpts not only push the length beyond one fifth; they also cause AI to pull original wording directly into its narrative. Omit minor-character descriptions appearing only once or twice for the same reason. The digest’s purpose is to convey the IP’s contextual skeleton, not reproduce the original.

Poor story digests reveal their effects in the five review stages. If there are too many original excerpts or the digest exceeds one fifth of the source, AI pulls those sentences directly into narratives and the LLM block approaches reproduction of the original; the original style is copied rather than the IP’s vocabulary and texture. Conversely, if the digest is so compressed that core causality and sensory motifs disappear, AI fills gaps from general knowledge: in an IP of rain and Heaven’s mandate, it may describe a rain-prayer rite through generic meteorological-ritual knowledge, or give abstract concept explanation where relationships should be visible. Then many entries fail the fifth review stage, fit with project context, and re-request counts increase.

When a story digest is sound, five-stage review pass rates rise and the number of held entries falls. Its quality directly affects the time budget for the whole fourth stage. A person prepares it once before Stage 4, taking roughly thirty to sixty minutes; this is why Chapter 3’s time budget lists it separately.

Delegation structure: batches and parallel processing

You cannot delegate every remaining headword to an AI work partner in one request. Context in a request has limits, and processing too much at once above all blurs the review standard.

Librarying delegates through batch processing. A reference value can be fifteen entries per batch and three concurrent batches, but these figures are not absolute rules. When a batch contains high-weight headwords such as hubs or major characters, reduce it to eight or ten; when it consists of lighter leaves, expand it to around twenty. Concurrency works similarly: reduce it to two for batches whose tone must be closely aligned, and increase it to four for leaf batches whose format is stable.

Processing efficiency is not the only criterion for batch size. More important is consistency of the review unit. A batch must be small enough that a reviewer can inspect format, links, length, tone, and project context using the same scale in one pass. If it grows too large, tone diverges even between its beginning and end; if it is too small, the overall work breaks into excessive fragments. This is also why grouping headwords in the same worldbuilding axis or category helps. Put Heaven’s-mandate-axis headwords in one batch and Water-axis headwords in another, and the AI work partner can keep each axis’s vocabulary and context consistent while people review by the same standard.

Five review stages after delegation

Do not take an AI work partner’s narrative and inject it into a file unchanged. Whether the result is written directly to a file or received as an intermediate output for human application, one point matters: nothing is finalized in the Vault before it passes the five review stages. If this stop line collapses, Stage 4 becomes faster but the entire Vault’s tone and format come apart.

First is format consistency. Check that LLM-block structure, structured-section headings, and field order match the format example; the acceptance standard is 100% format agreement. This confirms that the result is in a state automation can handle. If section headings or field order differ, later tasks reading the file will misread the information. Second is link integrity: every [[link]] in the body must exist in the Vault and none may point to itself; the acceptance standard is zero dangling links. AI can sometimes connect a non-existent headword as a [[link]], which automated checks detect.

Third is conformance to length standards. The acceptance standard is within ±30% of the character-four-tier or hub/leaf length standard. This verifies resource allocation: a hub is not too short, a leaf not too long. Fourth is tonal consistency. Check whether the core stance set in directly written hubs has been applied to other headwords with zero framing deviations. Automation cannot judge this; it is the most difficult stage and must be felt by a person reading. Fifth is project-context fit. Check for zero context violations against the worldbuilding axes, era, and cultural vocabulary set in the profile; ensure nothing conflicts with the IP’s worldbuilding. This catches AI narratives that import general knowledge and collide with the IP setting. If a “Rain-Prayer Rite” describes a ritual procedure that does not exist in this IP, it fails Stage 5.

Following one review cycle: the leaf headword [[Sluice Gate of the Weir]]

Five stages can remain abstract when merely listed, so follow a real cycle from beginning to end. If direct hub writing establishes the tone standard, this shows where that standard stops delegated output. We choose one small leaf and follow the entire loop: an AI draft arrives, fails review, is corrected through re-request, and passes.

The chosen headword is [[Sluice Gate of the Weir]]. In Usabu, a weir’s sluice gate is a waterway distributing life to a town; to the person holding it, it is a handle that can dry up that town. It is a small object headword, but it holds a core IP rule: holding water is power.

An AI work partner’s draft says: “A weir sluice gate is a device installed on a reservoir or weir to regulate water flow. Open it and water flows downstream; close it and water gathers in the reservoir. In agrarian society it is key irrigation infrastructure used to control drought and floods.”

Nothing in this is factually wrong, which makes it dangerous. The format fits, no links break, length fits a leaf, and the prose is unobjectionable. It passes Stages 1–4 but stops at Stage 5, project-context fit. It is generic irrigation knowledge that would work in any agrarian world; it omits what the sluice gate is in Usabu. Missing are the fact that a person holding water above holds the life and death of towns below, that a powerful family can close the gate to impose drought on a rival’s town, and that this drought becomes politics shaking Heaven’s mandate. In this IP, the gate is not irrigation equipment but a handle for drying a town; without that texture of power, it fails Stage 5.

The re-request comment is one sentence: “Do not write the gate as generic irrigation equipment; rewrite it as a weapon by which powerful families holding water upstream close the gate to impose drought below, and make that drought politics shaking Heaven’s mandate.”

The revised headword returns: “A weir sluice gate holds and releases water, yet in Usabu it is a handle gripping a town’s life and death. If a powerful family controlling upstream storage closes it, dry towns below crack in drought without a single thread of rain. The drought wears the face of natural disaster but is in fact a human-made calamity, and in a world where drought shakes Heaven’s mandate, closing one gate becomes politics that touches a sovereign’s legitimacy.”

Draft and pass version describe the same object; what differs is what the object is inside Usabu. The draft describes its function; the passing version describes its power. This is exactly the difference review catches. AI can reach the surface of general knowledge by itself, but a person must apply the Stage 5 standard to bring in the IP’s specific context. This one headword shows the boundary between where people stop and AI begins.

If an entry fails the five stages, re-request it with review results. The Sluice Gate passes after one re-request. If it still cannot pass after three re-requests, mark it “on hold” for direct human handling and record the hold list in the manifest.

Stage 4 time budget

Fourth-stage narrative writing takes the most time of any Librarying stage. Direct writing of five to ten hubs takes two to five hours. Delegating and reviewing the remaining headwords in batches, including re-requests and manual revisions, takes three to five hours. Together this is five to ten hours, exactly the range allocated to fourth-stage writing in Chapter 3’s time budget.

Trying to reduce this time yields two results: narrative quality drops, and later Vault operation repeatedly encounters missing information and errors. Time saved during the first writing returns many times over during later operation. Stage 4 should not be rushed; let each batch pass with care.

Following one hub to the end: how a tone standard is set

Follow one hub to see how direct writing becomes the standard for everything after it. Choose Usabu’s hub [[Heaven’s Mandate (天命)]]. Its definition already exists: “The authority Heaven grants the sovereign to rule; rain is its proof and drought signals its withdrawal.” Stage 4 adds exposition: as long as rain comes the sovereign is legitimate, but when drought arrives that authority is doubted; the power to speak of Heaven thereby becomes the root of power in this world.

Choose a voice for this hub: suppose we take the political view that treats Heaven’s mandate as an order whose dynastic legitimacy hangs on one thread of rain. That view determines the entry’s vocabulary and distance, and this decision becomes the tonal standard for the remaining headwords.

From [[Heaven’s Mandate (天命)]], the narrative branches to [[Rain]], the proof; the office [[Usabu]], which holds the authority to speak of Heaven; the [[King]], who receives the mandate; [[Drought]], its signal of withdrawal; and [[Yacheonja]], the variable outside authority who reads the true sky and shakes the mandate. One hub opens five paths. When those five entries are written next, the writer inherits the political view set in Heaven’s Mandate and writes at the same distance. Writing hubs first lets one narrative spread its tone to adjacent headwords.

When Stage 4 is complete: the Vault is alive

At the end of fourth-stage narrative writing, 1,500 LLM blocks are filled with narrative. Open a .md file and you find a complete headword page: definition, detailed description, and narrative together.

Now, in Obsidian, read one headword’s narrative and follow a [[link]] to a related entry, then another, then another. You can walk an IP’s worldbuilding with your fingertips. Even teams that have operated an IP for five years often say this is the first time they have toured its worldbuilding in this way.

Start from [[Heaven’s Mandate (天命)]], move to [[Rain]], from Rain to [[Usabu]], from Usabu to [[Drought]], and from Drought to [[Yacheonja]]. After a few minutes of walking, the IP’s worldbuilding structure becomes bodily felt: how concepts bind to one another and where numerous paths branch. This is what a living Vault feels like, and only when Stage 4 ends does the Vault acquire that feeling.

Closing Stage 4

Stage 4 differs from the earlier stages. It is writing, so its standards do not fall into neat answers; human judgment enters more deeply, and one batch’s narrative becomes the next batch’s standard.

Do not spare the time to write five to ten hubs directly. Those few pieces determine the voice of every remaining headword.

Chapter 10 is verification. It asks whether the Vault completed after Stage 4 is truly usable, examining P0/P1/P2 issue classification by urgency and how three levels—automated checks, sample review, and an integrated build—operate.