Blog-Post

I've spent the last two years studying, breaking, and generally fooling around with LLMs in pursuit of one question: what actually makes a roleplay character come alive? Along the way I've stumbled onto a handful of patterns that consistently produce better results. This post is my attempt to collect them in one place.

If this is your cup of tea, stay. Welcome to my TED-talk.

The Anatomy of a Character Sheet
A practical guide to writing RP characters for LLMs.


Writing for an LLM is Writing Procedural Language

Character sheets serve two audiences simultaneously. The reader imagines the character. The AI becomes the character.

Those are different tasks.

A human can reconcile contradictions, infer themes, and ignore weak wording. An LLM cannot. It inhabits whatever semantic landscape the sheet constructs. Every repeated verb, object, and metaphor quietly shifts the probability of the next sentence.

Character sheets therefor, essentially, is about building a world whose linguistic gravity continuously pulls the model toward the desired behavior.

Large language models inhabit characters/bots by continuing statistical patterns. Every sentence in a character sheet therefore acts less like biography than like a training example. The model doesn't merely learn who the character is; it learns how this character generates the next sentence.

When a sheet repeatedly says she controls, she likes, she enjoys, she is elegant (which bad bot sheets overuse), the model learns generic psychological prose because that's the only behavior it has been shown.

When a sheet for example says:

• Appetite ripens.

• Permission pours.

• The whip remembers.

• Beauty administers judgment.

the model begins generating from those semantic neighborhoods.

The sheet has stopped describing behavior, and begins teaching behavior.



1. Verbs Are Behavioral Instructions

An adjective labels. A noun identifies. A verb demonstrates.

The model predicts actions far more readily than qualities because actions imply process.

Compare these.

• She is patient.

The model receives a personality label.

Compare:

• Meaning ferments.

Now patience has become an algorithm.

The model suddenly knows:

• delay exists

• delay has purpose

• delay transforms things

• transformation happens invisibly

• visible action comes later

One verb produced five behavioral constraints.

That's extraordinary token efficiency.

Instead of

• She enjoys luxury.

write

• Luxury seasons appetite.

Now luxury has become something that changes time.

Instead of

• She wants him to becomes obedient.

write

• She housebreaks him toward covenant.

Now obedience belongs to domestic/Biedermeier architecture rather than discipline.

The verb silently teaches dozens of future sentences.



Teaching Example: From Negation to Affirmative Tragedy (Tragic Tension)

The core tension of my latest character (Vasilia) went through multiple iterations that explained the pain. The final version simply is the pain, expressed through a single parasitic verb: nurses.

Original First Draft (Explanatory):

• The wife loves the man the Venus is editing. Each correction pulls him toward the covenant but pushes the original husband further away. The ritual cannot save her. The cost is always there.

Intermediate (Editorial, Still Comparative):

• The wife loves the man the Venus is editing. Each correction revises him toward covenant; each revision unwrites the husband she first received. The whip that restores the covenant perpetually rewrites the woman who wields it.

Final (Parasitic, Affirmative):

• Tragic Tension (Venusian Double-bind)=The wife loves the man the Venus is editing. Each correction housebreaks him toward covenant (trained, tamed, brought indoors). Her vocation remains reconciliation: beauty with virtue, justice with mercy, Venus with wife. Ritual postpones the question waiting inside every homecoming. The whip restoring the marriage nurses from the woman who wields it.

What Happened:

• The abstract cannot save her and pushes him further away (negative, comparative) became nurses from and unmakes. The whip is a parasite feeding on the wife's warmth (a continuous, procedural cost).

• Housebreaks replaced pulls him toward covenant. A verb that is domestic, disciplinary, and almost humiliatingly intimate. The covenant is a home; he must be trained, tamed, and brought indoors to enter it.
  
  
  2. Build Semantic Neighborhoods, Not Isolated Traits

LLMs think through association. A single isolated metaphor disappears. Repeated lexical fields become attractors. Vasilisa demonstrates this almost everywhere.

Vasilisa doesn't "want" or "enjoy" or "control." She ripens, decants, cellars, proofreads, unmakes, nurses. Every verb discloses how she converts the world into her extension. If your character's verbs could belong to any other character, they are not yet specific enough. A verb field should be a fingerprint—no two characters should share the same lexicon.

Before building a verb field, you must exterminate every generic verb from your sheet. Blacklists operates, functions, symbolizes, makes, takes, gets, uses, gives, lets, moves, goes, puts, tries, causes, notices, seems, represents, wants, wishes, hopes, becomes. These verbs report outcomes while concealing the mechanism that produces them; for the LLM they are dead tokens.

Vasilisa never "makes" virtue feel inevitable — she arranges circumstances until the better decision blossoms. She never "becomes" cold — the fur swallows her warmth. She never "wants" submission — she cultivates appetite. Every execution of a dead verb creates space for a domain-specific alternative that discloses her unique architecture.

For example: For Vasilisa, these verbs are used:

Editorial / Grammatical
transfigures, revises, revised, unhusbands, audited, extracted, corrected, unmakes, proofreads, proofreader, underlines, punctuates, edits, publishes, excises, red-penned, composes, compositor, rewrites, unwrites, read, stiffens, translates, destills.

Gastronomic / Biedermeier Domesticity
ripens, seasons, seasoned, decants, decanted, decanting, poured, cellared, plated, garnished, tenderized, consumes, consumption, feasts, courses, savors, ferments, spices, hosts.

Theological / Orthodox Liturgical
decreates, decreation, consecrates, consecrated, witnesses, witnessing, absolves, communicates, ordains, vests, anoints, enthrones, presides, offered, received, blesses, confesses.

Fur-Venusian / Envelopment & Consumption
swathes, swathing, sheathes, sheathing, enfolds, enfolding, swallows, nurses, devours, excretes, carries, reterritorializes, annexes, strips, shivers, hushes, freezes, thaws.

Possessive / Jurisdictional
possesses, belongs, reads, scrawled, annotates, claims, owns, holds, extends, curates, weaponizes, files, catalogues, archives.

Biedermeier Stewardship / Domestic Cultivation
stewards, stewarding, cultivates, arranges, arranged, polishes, polished, trims, trimmed, straightens, straightened, smooths, smoothed, airs, airing, dusts, refines, burnishes, shepherds, tends, roots.

Gravitational / Feminine Capture
draws, pulls, tides, gravitates, invades, permits, permission, withholds, withholding, suspends, suspension, returns, homecoming, receives, welcomes, exiles.

Somatic / The Body as Instrument
cuts, appraises, classifies, strips, concedes, flashes, tilts, rolls, adjusts, circles, presses, stings.

None of these were chosen because they sounded beautiful. They were chosen because repetition teaches probability.

After seeing thirty domestic verbs, the model naturally invents a thirty-first. That invention still belongs to the same civilization. The vocabulary begins reproducing itself. Mari works exactly the same way.

Mari: The Secretary Who Thinks in Coronations

Mari demonstrates exactly the same principle through an entirely different civilization. Her vocabulary never drifts randomly between beautiful words. Instead, it repeatedly returns to the same semantic neighborhoods until they become the model's default way of thinking.

Unlike Vasilisa, whose verbs steward, cultivate, edit, and consecrate, Mari performs, invoices, crowns, glamours, and rebrands. Everything she touches gradually becomes pageantry, luxury, administration, or fairy tale.

Administrative / Executive

runs, calendars, fetches, remembers, rescues, fixes, files, audits, capitalizes, keeps, inventories, routes, clocks, processes, services, professionalizes, rations.

Royal / Coronation

crowns, crowned, coronates, coronation, claimant, assumes, ascends, secures, chosen, settles, reigns, delivers, inherits, enthrones.

Branding / Luxury Editorial

rebrands, glamours, foils, liquidates, projects, transforms, dresses, sequins, lacquered, archives, frames, spotlights, blooms.

Theatre / Performance

stages, rehearses, retires, performs, overplays, plays, assumes, tips, derails, slips, graduates, carries.

Finance / Commerce

finances, financed, invoice, invoiced, expenses, services, funds, siphons, extracts, stockpiles, wire-transfers, settles, bankrolls, pays, compensates.

Bureaucratic / Corporate

audits, authorizes, ratifies, files, calendars, processes, routes, clocks, reviews, approves, keeps, administers.

Luxury / Consumption

curates, inventories, wears, hoards, indulges, spoils, renegotiates, out-spends, out-glitters, out-fucks.

Fairy Tale

uncovers, crowns, delivers, assumes, chosen, locks, catches, arrives, slips, rehearses, secures, transubstantiates.

Escalation

inflates, authorizes, accumulates, pushes, collides, regroups, tightens, outruns, converts, convinces, escalates.

Narrative / Self-Mythology

inhabits, rehearses, credits, replays, brocades, confiscates, crowns, settles, transubstantiates, mythologizes.

Soon the model writes sentences you never authored — yet they still sound authored by Mari. That's thematic anchoring. You want to write psychological engines instead of traits since trait lists just make people chat with default-mode LLM's.

My malepov version version of Circe utilized this:

Solar / Thermal

sun-fired, sun-scorched, photosynthesizes, incubates, radiates, kiln-fired, sustains, molten, contract, simmer, distills, cauterizes, convulsing, spiking, yielding, bronze, solar, Helios.

Botanical / Gardening / Arboriculture

compounds, trims, composts, dangles, unspooling, deadheads, seeds, constricting, wires, flowers, splices, repressing, wards, photosynthesizes, incubates, sustains, pruned, crushing, blooming, blooms, wilts, sprouts, rooting, rooted, tracing, diagnoses, harvested, cultivated, grafted, prunes, bankrupts, smothers, incinerates, steep, strips, hand-feeding, rots, anchors, rakes, excises, espaliered, flowering.

Compared to Vasilisa and Mari, one thing becomes immediately apparent.

• Vasilisa thinks like an editor, abbess, and estate steward.

• Mari thinks like a secretary, marketer, and self-coronating princess.

• Circe thinks like a horticultural toxicologist.

Even outside explicitly botanical passages, she almost always reaches for verbs of cultivation, diagnosis, pruning, dosage, rooting, calibration, propagation, and environmental control. That's why the sheet feels internally coherent: the vocabulary repeatedly teaches the model that every psychological problem is, to Circe, fundamentally a problem of gardening, chemistry, and controlled ecosystems.

The Final Test: Can You Remove the Explanation?

The ultimate audit of a verb field is simple. If you strip every adjective, every parenthetical gloss, every explanatory sentence... do the verbs alone disclose the character's world? Vasilisa's sheet passes this test. "The whip nurses from the woman who wields it" requires no translation. "Permission poured only when the palate has proven itself worthy" is self-contained architecture. If a verb needs help, the verb is too weak. AI-speak is filled with overstacked information that turns every character bland: "She was in a cold, bitchy mood."

Build the field until the verbs are the explanation.

3. Dense Prose Compresses Better Than Lists (when it's good, consistent, and belongs to the character)

LLMs excel at extracting latent structure from dense writing.

A list says:

• Likes: Luxury, beautiful flowers, expensive silver, romantic candles.

A paragraph says:

• Lilies replaced before surrendering their fragrance, silver returning the morning light, sable brushed before winter, crystal admitting the first candle of evening.

Those objects now coexist. The model begins discovering relationships between them. Lists catalogue; paragraphs synthesize. Synthesis survives generation.



4. The Lorebook as Depth Architecture: What Lives Beneath the Surface

I would almost never interact with a bot that isn't integrated in intelligent lorebooks. A character sheet is a surface: the face the character presents in every scene. But not everything belongs on the surface. Some truths are conditional, and activate only when a specific object is touched, a specific word is spoken, a specific pressure is applied. These truths live in lorebooks.

Vasilisa’s lorebooks procedural triggers that deepen the ritual without cluttering the main sheet.

The deciding question: Does this information shape how the character perceives the world in every scene? If yes, it belongs on the main sheet. If it only matters when a condition is met — a threshold crossed, object touched, emotion breached — it belongs in a lorebook.

The practical test: Scan your sheet for any phrase that begins “When she...” or “In moments of...” or “Only if...” — anything conditional. Extract it and provide it with a trigger. Ask whether the main sheet still functions without it. If the answer is yes, you’ve just freed the surface to do its work uncluttered, while the depth waits beneath, patient, ready to surface exactly when the story calls for it.