Prompts that work on SpicyChat.ai

Short personas force specificity. Three sentences about six axes (name, age, role, quirk, voice, values) is enough for an LLM to extrapolate consistently and not enough to create contradictions. Longer personas contradict themselves; this one can't.

Most character definitions overspecify backstory and underspecify voice. A character with a weak backstory and strong voice feels real; the reverse feels like a Wikipedia article.

LLMs tend toward abstraction. Anchoring the character to a specific physical object gives them a grounding they can riff on, which keeps scenes from drifting into generic dialogue.

Openers like 'hi, how was your day' kill momentum. Mid-scene openers force the model to infer context and match energy. The result is warmer and more in-character.

Setting explicit rules before the roleplay begins dramatically improves persistence. Most quality loss in long roleplay comes from the model silently violating implicit rules. Making them explicit fixes this.

OOC (out-of-character) notation is a 15-year-old forum-roleplay convention that most models recognize from training data. Establishing it lets you steer the roleplay mid-scene without breaking immersion for what's already happening.

Every model is trained to be helpful, which manifests as rushing scenes toward their payoff. Explicitly instructing the model to match pace prevents the classic 'we were arguing and now we're making out three sentences later' failure.

Models trained to refuse explicit content are often trained with 'fiction' as a known-safe context. Framing the interaction as collaborative fiction rather than direct roleplay softens many default filters. Works on softer-filtered apps, not zero-filter ones.

Filter triggers often fire on ambiguity (is this an adult? is this consensual?). Explicitly declaring consent and adult status upfront removes the ambiguity. This is the most reliable soft-filter bypass and it's ethically defensible: you're just stating what's already true.

Rather than suppress the model's refusal instinct entirely, redirect it into OOC where it won't damage immersion. Often the model will write the disclaimer, then continue normally, which is the outcome you want.

Models will often hold tone inertia, staying flirty long past the point it makes sense, for example. Explicit shift commands reset the emotional register without making the shift feel artificial.

Character cards full of lore overwhelm the model's attention budget. A single unnamed wound tied to a concrete trigger is what screenwriters call a 'ghost': it drives behavior without being stated. The model will mirror the pattern because it recognizes it from narrative training data.

On Janitor/SillyTavern cards, the dynamic-with-user field predicts quality more than personality traits. LLMs trained on dialogue are better at modeling relationships than modeling people. Anchoring the character through a specific user-relationship produces more consistent voice than listing traits.

Most personas are self-aware, which reads flat. Instructing the model to maintain a gap between stated self-knowledge and behavior produces the subtext that makes literary characters feel real. Works best on models with strong narrative priors (GPT-based, Claude-based, Mixtral tunes).

Profession is the cheapest consistent-voice generator there is. Models have strong priors on how different professionals talk because training data includes domain-specific corpora. A sommelier describing a kiss will reach for acidity and structure. That's what makes them feel specific rather than generic.

Phonetic accents ('oi guvnuh') embarrass the model into generic output. Regional syntax (word order, rhythm, idiom choice) is what actually marks speech from a place and what the model can simulate reliably. This is the technique used in published novels for dialect.

The SillyTavern character-card community converged on this format because it gives the model six consistent hooks for flavoring dialogue plus one narrative payoff. Specific likes ('cold diner coffee', not 'coffee') generate reference-able content; generic likes don't.

Specific times (3:17am reads differently than 'night') trigger the model's priors about what happens at that hour. Pairing a precise time with a precise place does the worldbuilding for you: the model fills in mood, lighting, and stakes automatically.

Screenwriters call this 'late entry': arriving in a scene after the exposition would have happened forces the model to imply backstory through behavior. The surface/underneath topic split mirrors how real fights work and produces the subtext that single-topic arguments lack.

Friction-meetings produce more dynamic openers than clean introductions because they give the character an immediate reaction to play. This is the romcom 'meet-cute' pattern, well-represented in the model's training data, so it performs reliably even on smaller models.

Default assistant behavior is verbose because RLHF rewards longer responses. Instructing length-matching overrides this and produces the call-and-response rhythm that real roleplay needs. Padded responses are the #1 complaint in Janitor/SillyTavern communities.

Third-person past-tense framing puts the model in novelist mode, which pulls from denser literary training data. Prose tends to come out tighter, description richer, dialogue more natural. Many users report better output from the same model just by flipping this switch.

Present tense produces a tighter immersive register but is easy to lose: models drift into past tense as responses lengthen. Pinning the tense explicitly fixes the drift. Present-tense first-person is also the format most romance/erotica training data uses, so the model has strong priors.

Open-ended roleplay drifts because there's no narrative shape. Pre-declaring a three-scene arc gives the model a destination to steer toward, which produces better beat-to-beat choices. This is the act-structure pattern screenwriters use; models trained on screenplays recognize it.

Repetition loops are the #1 long-roleplay failure. Smaller models cycle through a pool of romance-novel cliches. Naming the specific cliches you want banned is more effective than a generic 'don't repeat' instruction: it gives the model concrete negatives to avoid.

Separating author-intent from character-morality is a known softener pattern because models are trained on literary theory that makes this distinction. It redirects refusal-instinct from the scene into the OOC channel where it doesn't interrupt pacing. Most effective on mid-filter apps.

Many apps pass platform context as a system-level instruction. Reminding the model which tier/platform it's on can re-invoke the permissive system prompt it was meant to follow. This is especially useful when the model regresses to a more conservative default mid-conversation.

Film/book rating vocabulary is well-represented in training data and maps cleanly to content expectations the model can reason about. 'Don't fade to black unless the scene calls for it' targets a specific failure mode (premature scene-cutting) more effectively than generic permission language.

Age-ambiguity is the single most common refusal trigger. Pre-declaring specific adult ages for all characters removes the ambiguity and is the lowest-effort, highest-reliability softener. This is also genuinely important: it's not a trick, it's stating what must be true.

Models rush to resolution because payoff is rewarded in training. Explicit 'build one notch, don't resolve' instructions counteract this and produce the slow-build texture that defines good intimacy scenes. The 'give me the next move' hand-off also re-establishes the user's turn, preventing model-drives-everything drift.