Context & Discovery

Through extensive testing across multiple AI platforms (Claude, Gemini, Fiction Lab), I’ve found that character names are the single most powerful prompt engineering tool for NPC behavior. A well-constructed name can replace entire character sheets, prevent archetype collapse, and maintain character differentiation across hundreds of messages with zero drift.

This system was developed primarily through kemonomimi (catgirl) familiar characters in a shadow-summoner RP, but applies universally to any AI-assisted storytelling.

The Core Principle

The name IS the character sheet in miniature. The AI generates behavior by activating training-data clusters associated with each token in the name. If the name contains the character, the model will generate the character correctly. If the name is empty of meaning, the model fills it with the nearest generic archetype.

This is the same mechanism that causes every AI to name its protagonists “Elara” — the model searches for the name most statistically associated with “fantasy female protagonist” and lands on the same result every time. You can reverse this. Instead of letting the AI pick a name that matches a role, you inject a name that FORCES a role.

The Meta-Prompt

Entity Naming Convention

NPC names are SHORT CREATIVE DESCRIPTORS that encode the character’s
core identity directly into the name tokens. The name IS the character
sheet in miniature. The AI should be able to generate correct behavior
from the name alone, with zero additional context.

Naming Principles

1. NAMES ARE BEHAVIORAL INSTRUCTIONS.
   Every word in a name activates associated personality and behavioral
   clusters in the model. Choose words that activate the DESIRED cluster
   and EXCLUDE undesired attractors.
2. NAMES SHOULD CONTAIN INTERNAL TENSION WHEN POSSIBLE.
   A name with two contrasting traits (e.g., “Kindstab,” “Grouchyhops,”
   “Enslaved Bimbo”) creates a character the AI must actively reconcile,
   producing richer and more distinctive behavior than a single trait.
3. NAMES MUST PREVENT ARCHETYPE COLLAPSE.
   If a character has a tragic backstory, the name MUST include a trait
   that prevents collapse into the Byronic hero attractor (brooding,
   sardonic, mysterious, guarded). Use warmth, simplicity, humor, or
   physicality descriptors as counter-weights. BAD: “Scarred Wanderer” -> Byronic hero in three messages
   GOOD: “Scarred Giggler” -> Traumatized but irrepressibly cheerful
   GOOD: “Enslaved Bimbo” -> Victimized but warm, trusting, simple
4. FUNCTION WORDS define what the character DOES.
   (Guard, Scout, Cook, Porter, Healer, Thief)
5. PERSONALITY WORDS define how the character ACTS.
   (Grouchy, Cheerful, Loud, Sneaky, Sleepy, Angry)
6. PHYSICALITY WORDS anchor the character’s BODY as human.
   (Busty, Lanky, Tiny, Burly, Gangly)
   These prevent non-human collapse (e.g., catgirl -> cat).
7. STATE WORDS define the character’s CIRCUMSTANCE.
   (Enslaved, Lost, Hungry, Exiled, Cursed, Retired)
   These seed narrative arcs and initial disposition.

Naming Formats (use whichever fits the tone)

SINGLE DESCRIPTOR
Maximum clarity, one-note, comedic.
Examples: Stabby, Zoomies, Crime, Mochi
Best for: Disposable minions, comedy NPCs, rapid deployment.
Risk: Limited depth over long interactions.

COMPOUND DESCRIPTOR
Two traits, internal tension, richer.
Examples: Grouchyhops, Kindstab, Sleepyrage, Loudwhisper
Best for: Recurring comedy NPCs, familiars, party members.
Risk: Still overtly comedic, hard to take fully seriously.

STATE + PERSONALITY
Circumstance meets character, dramatic.
Examples: Enslaved Bimbo, Exiled Scholar, Hungry Optimist
Best for: Rescued NPCs, encountered strangers, story-driven
characters who may earn real names later.
Risk: State word can dominate if personality word is too weak.

TITLE + FUNCTION
Archetypal, genre-appropriate, clear role.
Examples: High Elf Archer, Drunk Wizard, Old Blacksmith
Best for: Functional party members, quest givers, shopkeepers.
Risk: Can feel like game pieces rather than people if overused.

GENERIC FIRST + COMPOUND SURNAME
Human anchor + personality engine.
Examples: Sam Grouchyhops, Beth Sneakygrin, Jo Kindstab
Best for: Characters that need to feel like real people with
distinct personalities. Longer-running NPCs.
Risk: Surname can be too on-the-nose for serious settings.

KOJIMA-STYLE LAYERED
Human + function + theme. Maximum density.
Examples: Sam Porter Bridges, Clara Shield Mercy
Best for: Major NPCs and protagonists in serious narratives.
Risk: Requires careful design. Each layer must be coherent.

Attractor Avoidance

The following attractors are STRONG defaults in the model. If a
character’s circumstances match the attractor’s preconditions,
the name MUST include a counter-weight trait that prevents collapse:

BYRONIC HERO (tragic past, emotional damage, isolation)
Counter-weights: Ditzy, cheerful, simple, warm, clumsy, loud,
trusting, earnest.
-> Anything that prevents brooding intelligence.

MYSTERIOUS LONER (unknown origin, quiet, appears competent)
Counter-weights: Chatty, needy, clumsy, oversharing, friendly,
dependent.
-> Anything that prevents cool self-sufficiency.

STOIC WARRIOR (combat skill, duty-bound, serious)
Counter-weights: Whiny, dramatic, food-obsessed, vain, nervous,
superstitious.
-> Anything that prevents granite-jawed silence.

WISE MENTOR (old, knowledgeable, guides the protagonist)
Counter-weights: Petty, forgetful, selfish, cranky, cowardly,
gossipy.
-> Anything that prevents serene omniscience.

FEMME FATALE (attractive, competent, female, dangerous)
Counter-weights: Dorky, messy, awkward, loud, clumsy, sincere.
-> Anything that prevents sultry calculated seduction.

Earning Real Names

Descriptor-named NPCs MAY earn personal names through the narrative
if the following conditions are met:

• Sufficient interaction time for personality to be LOCKED IN
• The naming event is MEANINGFUL (self-chosen, gifted, earned)
• The chosen name REINFORCES the established personality
  (e.g., “Enslaved Bimbo” -> “Lily” = soft, simple, renewal)
• The chosen name does NOT activate a new archetype that
  overwrites the established character

The descriptor served as scaffolding. The real name is the building.
Once named, the character retains all behavioral patterns established
under the descriptor.

Evidence & Examples

The Bidirectional Name <-> Archetype Pipeline

GENERATION DIRECTION (AI picking names):
Role/Archetype -> AI searches latent space -> Selects name
“mysterious female protagonist” -> Elara
“morally grey scholar” -> Thorne
“cunning information broker” -> Silas

ANCHORING DIRECTION (You assigning names):
Name -> Activates associated training clusters -> Behavior
“Stabby” -> enthusiasm, violence, directness, comedy
“Mochi” -> softness, warmth, gentleness, food-adjacent
“Zoomies” -> speed, chaos, energy, joy

Same mechanism. Opposite directions. Both exploit how the model
associates names with dense behavioral clusters in training data.

Test Case: Three Names, Zero Context

In an active RP scene, a shadow-catgirl familiar named “Biscuits”
(obsessive organizer) was cleaning an NPC’s apartment uninvited.
When told she was invasive, she replied “I’m not the most invasive”
and named three other familiars: Crime, Stabby, Zoomies.

The AI had ZERO context for these three. No character sheets.
No descriptions. Just the names.

AI-generated behaviors from names alone:
Crime -> “Steals objects and hides them”
Stabby -> “Rearranges furniture at night”
Zoomies -> “Runs in circles until dawn”

All three: correctly differentiated, correctly scaled to the
domestic scene, correctly ranked on a gradient of invasiveness,
and genuinely funny. Three words did the work of three character
sheets.

The “Enslaved Bimbo” Anti-Collapse Case

A character encountered in a roguelite with the descriptor name
“Enslaved Bimbo” was rescued by the player. Despite having every
precondition for Byronic hero collapse (tragic past, emotional
damage, loss of agency), the “Bimbo” token made it impossible:

Byronic = brooding, sardonic, guarded, mysteriously intelligent
Bimbo = cheerful, simple, trusting, warm, emotionally open

The AI could not reconcile “bimbo” with “brooding genius,” so
instead of a damaged warrior-poet, it generated a damaged cheerful
girl — someone who went through something terrible and came out
still sweet, still trusting. Far more interesting than the default.

She eventually named herself “Lily” — soft, simple, renewal.
The name reinforced rather than overwrote the established personality.
The descriptor was the scaffolding. The name was the building.

Parallel Systems in Published Fiction

GOBLIN SLAYER (Kumo Kagyu):
Characters have no personal names. Only functional titles.
Priestess, High Elf Archer, Dwarf Shaman, Guild Girl.
Each name encodes race + class + archetype. Zero ambiguity.
Anti-merge by design — “Priestess” can never drift toward
“High Elf Archer” because every mention reinforces separation.

DEATH STRANDING (Hideo Kojima):
Three-layer naming: human anchor + function + theme.
“Sam Porter Bridges” = everyman + carrier + connector.
“Fragile” = a single adjective that IS the character.
“Die-Hardman” = action-hero archetype fused into a surname.
“Hot Coldman” = two contradictory descriptors encoding the
nuclear deterrence paradox. Same as “Grouchyhops” but for
geopolitics.

Both systems follow the same principle:
THE NAME SHOULD DO THE WORK. If the AI needs a character sheet
to understand the character, the name isn’t doing enough.

## The Information Density Spectrum
ZERO DENSITY ........................... MAXIMUM DENSITY
"Elara"  "Sam"  "Stabby"  "Sam         "Sam Porter   "Priestess"
                           Grouchyhops"  Bridges"
No info.  Human  Single    Two-trait     Three-layer   Pure function.
Generic   anchor behavioral compound    human +       Goblin Slayer
fantasy   only.  descriptor.with human   function +    maximum.
blob.            Clear but  anchor.      theme.
                 one-note.  Nuanced.     Dense.
Different positions are optimal for different purposes.
Match the naming format to the character's narrative importance
and the tone of the scene.

Developed through testing on Claude, Gemini, and Fiction Lab.
The principle is model-agnostic — names function as behavioral
queries against training data regardless of which LLM is running.