Chapter 13. What I Learned While Shaping 1,500: The Landscape of Practice

Some things can be learned in advance but understood only by doing them. Reading about nine classifications and knowing them intellectually is nothing like deciding the classification of one row at a time in a 5,000-row CSV. Explaining the orthogonality of worldbuilding axes is also a different experience from actually assigning axes to 1,500 headwords and repeatedly deciding, while refining the ecological-drama world Daesu’s Back, whether a headword belongs to Daesu or to beast-heart.

This chapter gathers what practice taught me. It is not a lesson in method so much as a retrospective from someone who has gone through it once.

First, the numbers. The first Librarying cycle for one IP created 1,847 headwords. Including empty stub files used to point aliases toward canonical headwords, the full Vault contained about 4,200 files. A human reversed 83 synonym candidates proposed by the AI work partner in the Stage 2 CSV; 127 headwords were confirmed as hubs; and Stage 4 writing produced 214 re-request cases.

How those numbers came about is what this chapter describes.

What practice teaches

The boundaries of the nine classifications are blurrier than expected. In verbal explanation, the boundaries seem clear: “Daesu’s step” is a device, “back keeper” is a character, and “the village beneath scale-shadow” is a place. But actual Stage 1 extraction produces far more ambiguous cases than expected. Is “a beast-heart bearer fallen into beast-heart” a character, a device, or mise-en-scène? Is “promise” a concept or a value? Is “the song of promise” an object or mise-en-scène? Is “the way of reading step and breath” an object or a device? Such questions arise dozens of times through extraction.

At first this ambiguity felt uncomfortable, and I worried that classification consistency would collapse. After several real cycles, however, I learned that spending too long on an ambiguous headword is the greater waste. Decide provisionally and move on; Stage 2 human review corrects misclassifications. Of 1,847 headwords, only 62 classifications changed in Stage 2—about 3.4 percent. Perfect Stage 1 classification might have reduced that number, but the time spent pursuing perfection would have been far larger. Practice makes the reason for setting the Stage 1 pass line at 60 percent tangible.

Worldbuilding axes are adjusted to the end. Axes are set during profiling, before Stage 1 extraction, by designing how headwords will be divided. One IP began with six: Daesu, back keeper, beast-heart, village, relationship, and general. As extraction proceeded, headwords appeared that fit none cleanly. Is “fetching water in the valley river” a back keeper or a village? Is “the old beast-heart bearer’s household” beast-heart or relationship? Is “the song of promise” back keeper or beast-heart?

The answer is either to add an axis or broaden an existing definition. Practice taught that broadening a definition is usually better than adding an axis, because every additional axis makes assignment harder. If the beast-heart axis is broadened to “every headword directly related to Daesu’s inner mind,” the old beast-heart bearer’s household fits naturally there. That cycle ultimately finished with seven axes: the original six plus promise/season. Starting at seven would have yielded more consistent classification, so more time in profiling is worthwhile.

The most time-consuming task in Stage 2 is deciding synonym bundles. The AI work partner rapidly proposes candidates, but reviewing each one to decide “are these really the same?” is harder than expected. IPs operated for five years or more often contain headwords whose meanings changed by era: the same name refers to a different thing in edition one than in edition three. For example, “beast-heart” may mean Daesu’s inner mind itself in edition one, but the ability of a person to feel that mind in edition three. Only someone who knows the IP’s history deeply can decide whether to combine or separate the two meanings. Of 83 AI synonym proposals reversed by a human in one IP, 41 were precisely such cases of meaning changing by period.

This is not unique to Librarying. Mystra, the magic goddess of Forgotten Realms, was one person in the first edition, but after a major world-changing event another person inherited the same divine title; the Forgotten Realms Wiki gives the two Mystras separate pages connected by redirects. Marvel Comics faces the same question whenever several generations inherit a title. In a long-running IP, notation and meaning rarely correspond neatly one to one.

Direct hub writing determines everything that follows. Chapter 9 explains why writing representative hubs directly matters, but practice makes the importance more urgent. In one IP’s first cycle I wrote twelve hubs directly, and ultimately had to rewrite three. The first three were written in an encyclopedia-like tone because I had not found the right voice; by the fifth, I realized that was wrong and revised the earlier entries.

The first sentence of the initially flawed “Last” headword read: “Last is the headword for the day when Daesu stops walking and lies down. This day arrives when the beast’s aging reaches its limit and affects all aspects of village life.” It explains the concept encyclopedically, but says nothing about how that day is experienced by a person in this IP.

After revision, the first sentence became: “Last is the day only one person carries in advance in a world where everyone believes they will last forever. While the village does not know that day, the beast-heart bearer bears its weight alone.” It first tells what the headword means inside this world, letting the reader feel the IP’s temperature even in a single entry. That was this IP’s narrative voice.

Comparing delegation after all twelve hubs were complete with delegation after only the first three were complete made the difference unmistakable. Re-request rate was 11.7 percent in the former; in a trial of the latter it rose to 31.2 percent. The quality of direct hub writing thus determines the quality of delegated narration, numerically as well as intuitively. Spending adequate time on hubs ultimately reduces the total Stage 4 writing time.

Without profiling, Stage 2 takes twice as long. Once I skipped profiling and started directly with Stage 1 extraction. When the result arrived, I had no clear sense of how to establish the axes; moving on anyway meant deciding each headword’s axis improvisationally, so the judgment at the thousandth headword contradicted the judgment at the first. After finishing Stage 2, I had to redesign the axes and reclassify every headword. Thirty to sixty minutes of profiling prevents hours of rework. The lesson is simple: the desire to start quickly repeatedly tempts us to skip profiling, but neither a new IP nor a first Librarying cycle should skip it.

Work takes 1.3 to 1.5 times longer than expected. Chapter 3 budgets 9–16 hours for a cycle, but one IP’s first cycle took 21.5 hours, about 1.4 times the estimate. Profiling took one extra hour, synonym-bundle decisions two, and revising the first three hubs one. The second IP cycle took 14 hours, about 1.05 times the estimate, because the first cycle’s lessons made it faster; later cycles stabilized at ten to twelve hours. Knowing in advance that the first cycle takes longer reduces surprise during the work.

Graph view reveals unexpected things. I initially thought Obsidian Graph View was merely attractive visualization, but it became an inspection tool. An isolated node led to a headword whose detailed description contained no [[links]]; its Stage 2 CSV description had omitted related-keyword links, an error hard to notice otherwise. It also exposed an unexpected hub candidate: when node size was set proportional to citation count, a headword absent from the hub list appeared unusually large. It proved to be a core concept repeatedly used in several character descriptions, whose importance profiling had missed.

Review units matter more than fixed numbers. I began by remembering figures like fifteen entries per Stage 4 batch and three simultaneous tasks. They are useful starting points, but after several cycles I found the more important question was whether one batch could be reviewed under one consistent standard. Fifteen entries mixing protagonists and core devices were too heavy, while twenty leaf objects passed lightly. Batch size and parallelism are not constants; they must change with the weight of headwords and the reviewer’s concentration.

Automation must remain as a contract, not merely code. Skeleton building and link injection seem finished once they are written and run. In practice, if the input, output, and pass conditions are not recorded, it is difficult to trust and reuse the same automation in the next cycle. Each automation needs its role, inputs, outputs, execution point, validation criteria, and the boundary requiring human judgment. Once this contract existed, an error no longer began with “the code is strange,” but with “which pass condition broke?” Recovery time fell sharply.

What differed from expectation

The AI work partner’s Stage 1 extraction accuracy was higher than expected. I expected 10–20 percent classification error; the actual rate was 5–7 percent. Characters, places, and objects were almost always correct. The difficult cases were abstract classifications such as devices, concepts, and values, whose error rate was three to four times higher than concrete classifications.

The difference in re-request frequency between delegated narration supplied with model hub narration and narration supplied without it was also larger than expected. Without models, re-requests were 31 percent; with them, 12 percent—almost a threefold difference. It was striking how directly the quality of direct hub writing affected delegated narration.

Graph View’s usefulness also exceeded expectations. Besides isolated nodes, coloring by worldbuilding axis made each axis’s hub role immediately visible. The graph made it possible to see that hubs on the Daesu axis had far more connections than headwords on other axes.

Backflow did not continue forever, contrary to expectation. At first I imagined that every short story would create endless new vocabulary and the Vault would only grow. Once the Vault reached several thousand headwords, however, backflow naturally slowed, and the Vault settled itself around a certain scale.

Each world seemed to have a fitting headword scale. Some worlds feel complete at a few hundred entries; others reach their form only at several thousand. The number cannot be fixed beforehand. In operation, some IPs produce a sense that the world is sufficiently filled and difficult to extend, while others still feel empty and continue backflow for much longer. Active backflow signals that a world is still growing; slowing backflow signals that it has found its scale.

What I would do differently

First, I would invest more time in profiling. The budget says thirty to sixty minutes, but for an unfamiliar IP one to two hours is better. Draft several axis schemes, apply them to a sample of fifty headwords, and choose the scheme that works most consistently; that investment accelerates the whole of Stage 2.

Second, I would not try to distinguish classifications too finely in Stage 1. Trying to settle whether every headword is an object or a device makes the work drag. Stage 1 should pass lightly, trusting Stage 2 to decide.

Third, I would move hub confirmation to immediately after Stage 2 consolidation. I first decided hubs after Stage 3 skeleton building, but automatic counts are already available in the Stage 2 CSV. Earlier hub confirmation makes Stage 3 and Stage 4 priorities clearer.

Fourth, I would write fuller story digests. A story digest is the condensed IP context given to the AI during Stage 4. Compressing it to one tenth of the source in the first cycle left the AI without enough context; expanding it to one fifth in the second cycle visibly reduced re-requests. Care spent on a story digest saves time across Stage 4.

Fifth, I would confidently defer P2 issues to the next cycle. In the first cycle, trying to resolve even P2 issues immediately repeatedly delayed handoff. P2 means “no impediment to Vault use”; it can be handled during the next incremental update.

The most important lesson practice gave me

Librarying is not the work of making a perfect Vault once. It is the operation of steadily updating a Vault that has crossed the pass line. The desire to make everything perfect in the first cycle stops the work itself.

Only after completing an IP’s first-cycle Vault did I feel that it was not an ending but a beginning. Forty-seven P2 issues awaited the next cycle, and eighty-three leaf headwords still lacked narration, yet the Vault had crossed its pass line and was ready for handoff. Short-form writing can now begin, and every incoming backflow will keep the Vault growing.

Starting at 60 percent and continuing to fill it without stopping: that is Librarying’s operational philosophy.

Chapter 14 is the final chapter. It gathers the core of what this text has tried to convey and suggests where to go after reading it.