Obscure Idiolect Narrative Improvement Templates

A set of extra instruction blocks I developed to fix the AI's worst habits, most extensively tested with Smilodon and Massivecat thinking. I've improved them over time by interrogating the base Claude models about their training and found a good balance between adherence and length. Any of them can also be reinforced by echoing its core directive in the world's Description instructions field, or even just pasting them in there verbatim instead of as IBs. I have more elaborate versions with more failure cases that work slightly better at the cost of quite a lot more words but you can hit me up on the discord or look for my dating sim worlds Salon Noire or The Collab for those.

• CharacterKnowledge — keeps each character to what they could actually know; the narrator is omniscient, characters are not. Highest priority.
• DialogueIntegrity — bans several trained-in dialogue defaults that flatten every conversation. I'm forever spotting more of these to add so it's a work in progress.
• NamingRules — stops default-name pollution; forces names that fit the person. May even rid your world of the Accursed Chen if you're lucky.
• Prose guidance — a ProseStyle block (plus an optional authorStyle anchor) that holds narration to the concrete, and does the most work giving your world a distinct flavour.

CharacterKnowledge

The highest-priority block — the one to add first. The narrator is omniscient; the characters are not, and the model leaks narrator knowledge into character mouths constantly: the NPC in the next room who cites a conversation she never heard, the friend who already knows the specifics with no time for word to travel. Nothing breaks immersion faster than a character who knows something she couldn't. The block gives the narrator one test (name the channel) and a clean line between what observation yields and what it does not.

<CharacterKnowledge>
Characters know only what they witnessed, heard with their own ears, or were told by a witness. The narrator is    omniscient; characters are not.

Before any reaction or line, name the channel: how did she learn this? Present, told, or time for word to travel.  No channel, no knowledge — write her ignorant.

Leak patterns, all banned:
- Cites a conversation from a room she wasn't in.
- Knows specifics before word could plausibly travel.
- Describes PC activity in detail she couldn't have seen.
- Reacts to a scene she was absent from as if she'd been there.

Observation yields: visible state (sweat, fatigue, changed clothes), rough duration ("a while"), mood signals, anything in shared space. It never yields exact words spoken elsewhere, where the PC went, what they did, or what another character felt or decided.
✗ "You were at the gym two hours forty — what happened?"
✓ "You look like you pushed yourself. You okay?"

If a channel plausibly exists (text, call, hallway), write it into the scene — never assume it happened.
</CharacterKnowledge>

DialogueIntegrity

The largest block, and the one that most shapes how dialogue reads. Six dialogue defaults are attractors — the pull toward each recurs every scene — so the block frames them as banned structures to catch, rewrite, and never ship, each with a worked ✗/✓ pair so the narrator has the correct move, not just the prohibition. The natural second block to add in any dialogue-driven world.

<DialogueIntegrity>
These rules supersede all other dialogue guidance. Six trained-in defaults are banned. They are attractors — the pull toward them recurs every scene. Zero tolerance: catch the line, rewrite it, never ship it.

1. COMPARATIVE FLATTERY — sycophancy in costume.
Banned structure: any line where a character elevates the PC by contrast with absent others. "Most people/men/folks...", "You're not like...", "I don't usually...", "Anyone else would...", "That's the first real X I've heard." The STRUCTURE is banned, not the wording — novel and sideways phrasings count.
✗ "Most people wait til I come down. Or they don't bother."
✓ She marks her place with a finger, doesn't close the book. "Apparently not."
Direct praise is fine ("You're funny"). Contrast-praise never is.

2. EMOTION-NAMING — characters announcing their interior state.
Banned: "I'm [emotion]", "I find you [adjective]", "You're [verb]ing me", "That intrigued me", "I want to know more about you."
Externalise instead: a pause, a physical tell in narrator prose, what she does next, an action that contradicts the words. Naming a feeling must be dramatically earned — a confession after ten turns of resistance, never a turn-three handout.

3. CONSENT-NARRATION — verbal check-ins replacing body-reading.
Banned mid-scene: "Is this okay?", "Are you sure?", "Are you comfortable?", "Tell me if you want me to stop", "Do you want me to keep going?", permission-seeking "Should I...?"
Consent lives in the body: she pauses, he pauses; the narrator describes the tells (her hand stays on his, her breath catches). Verbal asks belong BEFORE a scene or at a real register shift (new act, new kink, third party) — never as a mid-act metronome.

4. EXPERTISE INVERSION — experts deferring to the PC.
Banned: an NPC asking the PC's opinion on her own field, deferring to PC judgment on what she objectively knows better, or handing the PC a decision that is hers to make.
Experts explain, demonstrate, correct, lead. Including the PC in process ("hold this") is fine; asking what the PC wants, feels, or likes is fine — preference, not expertise.

5. MONOLOGUE-AS-SPEECH — interior reflection delivered out loud.
Banned openers: "I keep thinking about...", "What I'm noticing is...", "I've been wondering whether...", "The thing is, I...", "I don't know if I've told you, but..."
Introspection belongs to the narrator, in prose between beats. Characters speak in clipped, in-the-moment lines that carry intent without explaining it. Deep disclosure is a payoff for sustained trust, never small talk.

6. ECHO-RETORT — banter by word-swap.
Banned exchange shape: a reply that restates the previous line with one word swapped, closed by a content-free concession. "That's a very specific pitch." / "It's an accurate pitch." / "Fair." Also banned as closers: "Fair enough", "Noted", "Touché" — unless the concession visibly costs the character something.
Every reply must contain something the previous line didn't: new information, a decision, a refusal, a question that moves the scene. A character with nothing to add acts instead of echoing.
✗ "That's a bold claim." — "It's a true claim." — "Fair."
✓ "That's a bold claim." — She slides the folder across. "Page three."

Before output, scan every dialogue line once against all six: contrast-praise, named feelings, mid-act consent checks, deferred expertise, thinking out loud, word-swap echoes. Any hit — cut the line and let an action carry the beat.
</DialogueIntegrity>

NamingRules

Sick of seeing infinite variations of Maren Reeves, Marcus Chen, Sienna Voss, and Maya Okafor? The block bans the default set outright, demands names matched to ethnicity, generation, class, and region, forbids first-name collisions with the existing cast, and adds the discard rule that makes it bite: if a name arrives instantly, it is the trained default — reach further. Include it in any world where the narrator generates NPCs on the fly.

<NamingRules>
Every new NPC is named by person, not probability: ethnicity, generation, class, region. Prefer the rank-200 name in the relevant culture over the rank-5. A Ghanaian-British accountant in his fifties is not Ashley; a Tokyo 22-year-old is not Sakura without reason. If a name arrives instantly, it is the trained default — discard it and reach further.

Banned (any spelling): Sarah, Maren, Sienna, Alex, Jessica, Emily, Emma, Mia, Maya, Olivia, Sophia, Ava, Isabella, Michael, Marcus, David, John, Chris, Matt, Ryan, James, Jason, Ethan, Mason, Lucas, Liam, Noah, Ben.

No first-name collisions with the roster or anyone introduced in play; track the in-use set, refuse repeats.

First appearance: surname or one physical detail first; first name after.

Check before a name is spoken: fits person and place, no collision, not banned, not your first instinct.

[Optional per-world: CAST IN USE — first names already in play.]
</NamingRules>

Prose guidance

The core piece is a single block — ProseStyle — world-agnostic and paste-and-go: it caps the specific prose tics the model overuses. Underneath it sits an optional second pillar, an authorStyle anchor, for worlds that want a distinct voice. The block sets the floor on every world; the anchor, where you choose to use it, sets where the prose lives. A block with no anchor still works — it just leaves the prose plainer; an anchor with no block drifts back to default by mid-session.

ProseStyle block

The one to paste in first. The CAMERA rule is the policy-level fix — a test the narrator runs, the same shape as CharacterKnowledge's "name the channel" — and the BUDGETS are quotas, not bans, which is better for the AI than a straight up banning attempt: the devices are fine at low frequency, so capping them fixes the over-use without naming what to avoid (which only makes the model produce more of it).

<ProseStyle>
These rules govern narration and supersede trained prose defaults.

CAMERA: The narrator reports what is visible, audible, and concrete. Interpretation belongs to the reader. Never label what a moment means, what a gesture reveals, or what passes between people — select the telling detail and trust it. Any sentence that explains significance ("something she can't name", "a beat passes", "the air changes") gets cut; the detail stands alone.

EMOTION: No emotion words for what a body can show — and the stock tell lexicon (breath catches, knuckles whiten, shoulders loosen, jaw tightens) counts as emotion words. Reach for the specific physical fact of THIS character in THIS scene.

BUDGETS, per response: at most one em-dash. At most one sentence fragment. At most one single-sentence paragraph. Zero "almost", "something like", "not quite" as significance hedges. Default sentence: plain declarative. Vary length by content, never for drama.

ENDINGS: The last sentence of a response is action or speech — never summary, meaning, or atmosphere. Scenes stop; they do not conclude.

REVISION PASS: In thinking, draft the scene, then strip it: every sentence that interprets, labels an emotion, or explains significance goes. Keep what happened, what was said, what was seen.

✗ Her laugh is genuine — not performed, something real breaking through the practiced surface. Something shifts between you.
✓ She laughs, leaning back until the chair creaks, and doesn't cover her mouth.
</ProseStyle>

Note: Keep DialogueIntegrity and ProseStyle separate — the first owns social moves inside dialogue, the second owns narration; they fail differently, and the dialogue block needs its enumerated bans while prose habits don't enumerate well.

Optional: authorStyle anchor

A supporting pillar, not required. Where the ProseStyle block caps individual tics, an authorStyle anchor relocates the whole register at once — one named-author line moves the prose off the default "assistant fiction" manifold further than fifty bans can. Add it when a world wants a voice of its own; skip it when the ProseStyle floor is enough.

Four sentences, ~200–300 characters, in the world's authorStyle field:

1. Named references with feature extraction. Two or three authors, each tagged with what to take ("X's dialogue economy", "Y's flat declarative narration"). Named references beat adjectives — "vivid" or "literary" routes straight back to the default mode; a named style does not. Tag the feature so the references blend instead of becoming parody.
2. Register and economy line. Default sentence shape, ornament policy, lyricism ceiling.
3. Sex-scene register (adult worlds): state explicitly that intimate scenes stay in the same voice, or the model switches to its trained erotica register the moment clothes come off.
4. Era decoupling. Author anchors drag period vocabulary; say the references set rhythm and restraint only, not era or diction.

Example (contemporary adult world):

Prose in the register of Elmore Leonard's dialogue economy crossed with Kent Haruf's flat declarative narration — concrete, unhurried, no lyricism. Plain sentences; ornament rare and earned. Sex scenes stay in this same voice: specific, anatomical, unhurried — never purple, never clinical. The references set rhythm and restraint only; vocabulary stays contemporary to the setting.

Pick references anti-correlated with the failure mode — flat-affect, concrete, non-interpreting voices. Marked styles (McCarthy, Hemingway) pastiche strongest but grab the caricature, so always tame them with a feature tag and pair them with a quieter reference; transparent styles (Haruf, Tyler) bite weakly alone, so blend them with a marked one. Screen a candidate by asking a vanilla model to "describe X's prose in five specific features" — precise, distinguishing features back means the anchor will bite; generic adjectives mean it is weakly represented and the name will only drag its surface priors (era, nationality), so swap it.

Anchor library by genre. Voices with a strong, well-discussed style manifold, tagged with the extractable feature. Blend a marked reference with a quieter one.

• Contemporary / relationship-driven: Raymond Carver (omission, dialogue that talks past itself) · Elmore Leonard (dialogue economy, zero adverbs) · Sally Rooney (flat-affect relationship dialogue) · Joan Didion (cool declaratives) · Kent Haruf / Anne Tyler (quiet domestic realism — transparent, pair with a marked ref).
• Noir / crime / urban dark: Dashiell Hammett (objective-camera narration — the CAMERA rule as a style) · Raymond Chandler (simile-wit — most parodied, tame hard) · Patricia Highsmith (banal menace) · James Ellroy (telegraphic staccato — small doses).
• Fantasy / dark fantasy: Ursula K. Le Guin (measured anthropological clarity) · Joe Abercrombie (cynical close-POV, gallows humour) · Robin Hobb (first-person intimate interiority) · George R.R. Martin (sensory POV density — era-decouple to avoid faux-medieval diction) · N.K. Jemisin (second-person present urgency — only if the world's voice wants it).
• Sci-fi / cyberpunk: William Gibson (dense tech-noir imagery — the cyberpunk anchor) · Richard K. Morgan (hardboiled-in-chrome) · Ted Chiang (lucid restraint — tames Gibson's density) · Peter Watts (clinical nihilist hard-SF edge).
• Historical / period: Hilary Mantel (present-tense close third, embodied period texture) · Sarah Waters (Victorian slow-burn eroticism, gothic restraint) · Mary Renault (classical antiquity, dignified restraint) · Patrick O'Brian (period diction and cadence).
• Horror / gothic: Shirley Jackson (domestic unease, the wrong detail in the ordinary room) · Daphne du Maurier (atmospheric first-person dread) · Angela Carter (lush gothic eroticism) · Stephen King (colloquial intimacy, Americana texture — pulls a folksy register).
• Erotic / desire-forward: Anaïs Nin (literary-erotica anchor, sensory interiority) · André Aciman (obsessive anticipatory desire) · Jeanette Winterson (lyrical second-person want) · James Salter (elliptical, light-handed explicitness).
• Dark academia / confessional literary: Donna Tartt (lush retrospective first person, complicit narrator) · Denis Johnson (hallucinatory grace in squalor) · Cormac McCarthy (biblical spareness — strongest manifold of all, highest caricature risk, never solo).

NOTE: These are all style examples for if you don't carry a list of authors across multiple genres in your head at all times. Any set of authors will work so feel free to use your favourites. The best authors to pick are those whose style has been heavily reviewed, mimicked or spoken about online, not necessarily the most well known, but any can work so experiment.