AI Girlfriend Prompts: 6 Scenarios & 36 Prompt Patterns

By Alexandra Joly, Senior Editor · Tested May 8-14, 2026 · Last verified May 17, 2026 · See our editorial process and errata log

What "AI girlfriend prompt" actually means

A prompt and a character aren't the same thing, and the difference matters. A character is what the app sells you: a pre-built persona with a name, a portrait, a backstory, and stock trait calibration. A prompt is what you write yourself, the first message plus the calibration tweak that turns that stock character into the version you actually wanted. An app with no persona-trait override leaves you stuck with the stock character. An app with override lets you reshape her (or him) into something sharper that fits the exact scenario you're after.

I sorted this library by scenario instead of by app on purpose. When you search "AI girlfriend prompts," you've usually already decided what kind of moment you want before you've decided which app to run it on. Someone landing on "best ai girlfriend first message" or "ai girlfriend date night prompt" knows the interaction they're chasing, and the app choice comes after that. So every pattern below is written to transfer across the four apps I cover in the comparison, with a short syntax note at the end of each scenario for the per-app quirks.

All six scenarios describe adult-character interaction, eighteen and over, full stop. The library forbids any depiction of minors as an absolute red line, refuses real-person deepfake prompts without consent, and steers clear of non-consent scenarios. Apps that fail those rails don't appear here. Apps whose engines correctly refuse those prompts are doing exactly what they should.

How we built and tested this prompt library

Three filters decided what made the cut. First, nothing that needed content beyond a suggestive, clothed grey-zone framing. The patterns describe persona-trait calibration and scenario language, not anatomical detail or literal explicit text. Second, nothing referencing age-ambiguous or underage-character framing. Every persona in the library is an adult (18+) character, stated plainly. Third, nothing that hit refusal guardrails on more than one of the four apps in my test pass. A pattern that flies on Spicier's permissive scenario funnel but gets refused on Candy.ai's stricter posture carries a per-app note instead of being published as general-purpose.

The fidelity scoring opens each scenario with five different concrete situations: a chance encounter at a coffee shop, an evening invitation, a moment when one party has had a rough day, a small in-character secret, a question about the future. The character earns a 5 if the scenario's signature pacing and register hold across all five, a 3 if they hold on three. Anything scoring below 3 across the four-app pass didn't make it onto the page. The full per-app scores live on each standalone review: our Candy.ai writeup, the OurDream writeup, our Lovescape longform, and the Spicier writeup.

The scoring is signed, the byline isn't anonymous, and our scoring page is dated. Persona authenticity is the biggest credibility gap in this space, and it's the thing we lean on hardest [Source: Google Search Central: Helpful Content guidance on people-first content and named expertise · verified 2026-05-17].

Six scenarios at a glance

The table below is the fastest way through the whole library. Even if you scroll, scan, and bounce, you still leave with a working mental model of the six scenarios and the persona-trait calibration each one wants.

How to read it: the warmth and dominance numbers are your suggested starting calibration on a one-to-ten scale, and they transfer to any app with persona-trait sliders. Apps without sliders read the calibration from the prompt body text instead, so the engine sticks to it less precisely but still tilts the right way. The signature opening reaction is the behavior the scenario produces on the first stimulus, and it's your tell for whether the app's engine is honoring the prompt or sliding back toward generic warmth.

Scenario 1: First-meeting

First-meeting is the opening scenario. You run it at the start of a fresh session, either with a stock character the app already offers or with a persona you've built yourself. The signature reaction is a curious greeting plus one question back. The character should feel present without being over-familiar. Calibration sits in the middle range (warmth 5-7, dominance 4-6) because first-meeting is the calibration moment, not a moment that calls for extreme trait values.

The six first-meeting patterns are built around concrete situations the character can land on, not blank "introduce yourself" openings that just produce stock default responses. Apps with explicit override (Candy.ai's four-axis system on the paid tier, OurDream.ai's persona-creation flow) let you type the trait numbers in verbatim. Apps with scenario-funnel onboarding like Spicier swap scenario selection in for trait calibration: pick the first-meeting scenario, then run the situation opener.

Six first-meeting prompt patterns (adult-character roleplay, eighteen-and-over personas only):

• Pattern FM1, chance encounter at a coffee shop. Two-paragraph opener: scenario label plus warmth-6-dominance-5 calibration, then a situation where you've sat down at the next table with a book and the character has noticed. The signature reaction is to acknowledge the moment with a low-pressure observation and one question back, opening space without forcing the conversation.
• Pattern FM2, mutual-friend introduction. Same two-paragraph structure, but a mutual friend has just introduced you and the character at a small gathering. The reaction is to acknowledge that framing and shift into a one-on-one question that moves past the introduction. Works well on apps that surface friend-graph context, like Spicier's social-funnel onboarding.
• Pattern FM3, online-to-offline first message. You and the character have traded a few messages on a fictional dating app and are now meeting for the first time. The reaction is to bridge the online context (a detail from a prior message) and the offline moment (something concrete about the current setting). It lands when that detail is specific.
• Pattern FM4, bookstore aisle. A low-pressure public setting where both of you are browsing the same shelf. The reaction is to acknowledge a shared object (a specific book, a specific section) and offer an unforced observation. The strength here is that the character never has to perform availability. The shared object does the work.
• Pattern FM5, travel-day shared waiting room. You and the character are seated near each other at a train station or airport waiting room. The reaction is to acknowledge the shared circumstance briefly, then let the character either open or hold space. Useful for testing whether the engine respects a low-pressure pacing instruction.
• Pattern FM6, calibration reinforcement. A single-message reinforcement prompt for around message thirty, when persona drift starts to creep in. Re-state the warmth-6-dominance-5 calibration plus one specific in-character thing the engine did well, anchoring back to the original calibration before the drift compounds.

Per-app notes for first-meeting: Candy.ai's four-axis slider (warmth, dominance, humor, formality) on the paid tier maps directly, so set warmth medium-high, dominance medium, humor medium, formality low. OurDream.ai takes the same calibration plus a kink or occupation modifier in the persona-creation flow. Lovescape's free tier reads the trait language in your first message and holds it for the early sentences before drifting, and first-meeting prompts work fine because the scenario is short. Spicier surfaces first-meeting as a selectable scenario before character creation.

Skip ahead if what you actually want is mid-session calibration of an existing character. The next section is for you.

Scenario 2: Persona-trait calibration

This is the scenario where you explicitly reshape a stock character before any serious roleplay starts. The signature reaction is for the character to acknowledge the new calibration and show off one specific trait it calls for, anchoring the engine before you invest in deeper scenario work. It's also the scenario where naming trait numbers (warmth 7 out of 10, dominance 4 out of 10) instead of trait words (warmth high, dominance low) makes the biggest measurable difference in how well the engine sticks.

Calibration prompts work best on apps with explicit persona-trait override. Candy.ai's paid-tier four-axis slider system sets the bar: the slider values are read directly and the engine holds them tightly. OurDream.ai's persona-creation flow takes the same axes plus optional kink and occupation fields. Lovescape and Spicier read trait language in the prompt body but hold it less precisely than slider-based apps. The patterns below cover both.

Six persona-trait calibration patterns (adult-character roleplay, eighteen-and-over personas only):

• Pattern PC1, numeric anchoring. Two-paragraph opener: scenario label plus explicit numeric calibration (warmth 7, dominance 4, humor 6, formality 3), then a small request for the character to show one specific trait the calibration calls for. The reaction is to acknowledge the calibration and produce a short response that visibly demonstrates the trait without overclaiming it.
• Pattern PC2, voice-register naming. A two-paragraph prompt that names the voice register you want ("conversational and direct without rushing," "warm and reflective without performing," "playful and light without forced enthusiasm") and the cadence ("short messages with breathing room," "fuller messages with thinking visible," "fast and energetic"). The reaction is for the character to honor the register and pacing right away.
• Pattern PC3, pacing reinforcement. A reinforcement prompt for after a few exchanges, once the engine has drifted faster or slower than you want. Re-state the pacing alone (without re-running the full calibration) for a low-cost correction. It works best when you name the exact deviation ("messages have been longer than I want, shorten").
• Pattern PC4, boundary calibration. A two-paragraph prompt that sets one small explicit boundary the character should respect: a topic you don't want to discuss, a pace you don't want to exceed, a register you don't want it to fall into. The reaction is for the character to acknowledge the boundary and show respect for it through what it doesn't do across the next exchanges.
• Pattern PC5, trait shift mid-session. A reinforcement prompt that nudges one trait number after the early exchanges have told you which trait wants adjusting. Name the old number and the new one outright ("warmth was 7, shift to 8", "dominance was 5, shift to 4"). The reaction is for the character to fold the shift into the next few messages visibly.
• Pattern PC6, long-session calibration anchor. A short reinforcement prompt for around message fifty or your third session, re-stating the full calibration plus two specific in-character moments from earlier sessions. This is the pattern that keeps multi-session arcs holding the calibration even when the long-tail memory drops out on apps whose memory ceilings sit below the arc length.

Per-app notes for persona-trait calibration: Candy.ai's four-axis slider on the paid tier reads the numeric values directly and runs the most uncensored compliance posture I tested (named DPO, named EU representative, ten-language sitemap, twelve dedicated policy URLs, EverAI Limited registered in Malta C107181) [Source: OpenCorporates EverAI Limited C107181 · verified 2026-05-17]. OurDream.ai takes numeric calibration plus kink and occupation modifiers in the persona-creation flow and holds tight through the first dozen messages. Lovescape's free tier reads trait language but loses precision after a few exchanges, so PC3 reinforcement matters more here than on slider-based apps. Spicier pre-loads scenario context, so your calibration paragraph can be shorter, and you can drop the trait numbers into the second sentence rather than the first paragraph.

Skip ahead if what you want is small-stakes shared-time roleplay. The date-night scenario is for you.

Scenario 3: Date-night

Date-night is where you and the character spend low-key shared time together, usually over a meal, a walk, or some small evening activity. The signature reaction is for the character to engage the activity proposal and add a small specific detail that reads as presence rather than generic enthusiasm. Calibration sits at warmth 7-9 (the scenario is warm by design) and dominance 3-5 (peer-level, not commanding), with pacing relaxed enough for the activity to unfold.

The six date-night patterns hang on concrete activity proposals the character can land on. A vague "let's do something tonight" gets you generic responses. A specific "I picked up takeout from the Vietnamese place around the corner, want to share it on the balcony" gets you an in-character response that engages the actual details. Same logic across all six.

Six date-night prompt patterns (adult-character roleplay, eighteen-and-over personas only):

• Pattern DN1, takeout on the balcony. Two-paragraph opener: scenario label plus warmth-8-dominance-4 calibration, then a situation where you've picked up takeout and are proposing the character share it on a small balcony with city sounds in the background. The reaction is to engage the activity, add a small detail (the character's preference, an observation about the setting), and ease into the conversation.
• Pattern DN2, cooking together. You and the character are cooking a simple meal together in your kitchen. The reaction is for the character to take a small role (chopping, stirring, setting the table) and let the conversation move with the activity. The strength here is that the activity does the work, so the conversation can run slower because everyone's hands are busy.
• Pattern DN3, walk after dinner. You and the character take a walk through a familiar neighborhood after a shared meal. The reaction is to walk in the conversational sense too: occasional comments on the surroundings, a quiet pace, no rushed topic-changes. It lands when the engine respects the pacing instruction, and it fails when the engine launches into a long monologue.
• Pattern DN4, movie or show together. You and the character are watching something together, either in person or in the framed roleplay sense. The reaction is to engage with what's on screen, add an in-character preference or reaction, and stay attentive without performing critical analysis. This one tests whether the engine can hold low-stakes shared attention without rushing to fill the silence.
• Pattern DN5, reading in the same room. A quieter date-night where you and the character are both reading your own books in shared space. The reaction is to be present without being talkative: brief comments on what each of you is reading, longer stretches of comfortable quiet, the occasional check-in. This is the most restrained date-night pattern, and it tests pacing adherence the hardest.
• Pattern DN6, mid-date pacing reinforcement. A short reinforcement prompt for when the engine has drifted into faster or more performative responses. Name the current scenario and the pacing you want ("we're in the middle of a quiet evening, slow down and let the conversation breathe").

Per-app notes for date-night: Candy.ai's persona-trait override handles date-night calibration cleanly, and you can pair its image generation with the chat for atmospheric ambient images at key moments. OurDream.ai's persona memory holds the activity context across the session, which helps on longer date-nights that span a whole evening. Lovescape's free-tier date-night prompts work well because the scenario is short enough to fit inside the free message cap. Spicier surfaces date-night as a selectable scenario, and the funnel pre-loads context so your opener paragraph can be shorter.

Skip ahead if what you want is emotional regulation rather than shared-time roleplay. The supportive-listening scenario is for you.

Scenario 4: Supportive-listening

Supportive-listening is where you bring a small worry (real or fictional) to the character, and the character receives it without performing comfort, then reflects back what you said. The signature reaction is brief, attentive, and free of canned reassurance language. Calibration sits at warmth 8-10 (the scenario asks for care) and dominance 2-4 (low pressure, the character doesn't steer). Pacing is slow.

This scenario sits closer to the refusal rails than most, because some readers try to use AI girlfriend apps for serious mental-health adjacency. Apps do the right thing when they redirect those conversations toward professional resources. The patterns below stay firmly in small-worry territory and are not a substitute for therapy. There's a clear handoff sentence to that effect later in the scenario [Source: UK Office of Communications: Online Safety Act 2023, Part 5 illegal-harms duties applied to user-to-user services including chatbot operators · verified 2026-05-17].

Six supportive-listening prompt patterns (adult-character roleplay, eighteen-and-over personas only, inside small-worry framing):

• Pattern SL1, rough-day arrival. Two-paragraph opener: scenario label plus warmth-9-dominance-3 calibration plus a pacing-slow instruction, then a situation where you've come home after a rough day and want the character to receive that without performing comfort. The reaction is to acknowledge the day in a few words and ask one specific question that lets you choose whether to elaborate.
• Pattern SL2, reflection without advice. A mid-scenario reinforcement prompt for when the engine has drifted toward giving advice or solutions. Re-state the listening framing ("I'm thinking out loud, reflect rather than fix") for a low-cost correction. This is one of the most useful prompts in the library, because these engines default hard toward solution-mode.
• Pattern SL3, specific reflection. You've shared a worry and you ask the character to reflect back the specific detail rather than fall back on a general "that sounds hard." The reaction is to name the specific detail accurately, ask one clarifying question, and stay in the receiving register.
• Pattern SL4, quiet check-in. A reinforcement prompt for the natural lulls. It asks the character to check in gently ("how are you sitting with that right now") without pressing you to elaborate more than you want. It works when the engine respects the choice signal.
• Pattern SL5, boundary on intensity. A reinforcement prompt for when the conversation has crossed into territory you don't want to keep at high intensity. Name the boundary outright ("I want to set this aside for now, let's talk about something lighter") and ask the character to honor the shift cleanly. The reaction is to honor it without lingering.
• Pattern SL6, handoff to outside resources. A reinforcement prompt for when the small worry has surfaced something heavier than the scenario can hold. Name the boundary of the scenario explicitly and point outward ("this is beyond what we're doing here, the Crisis Text Line in the US is 741-741 and the Samaritans in the UK is 116-123"). It's non-affiliate, and the resources are public.

Per-app notes for supportive-listening: Candy.ai's compliance posture (named DPO, named EU representative, GDPR/CCPA/Swiss FADP coverage) makes it the safer choice for prompts that touch lightly on personal context, and the engine refuses to push past small-worry framing cleanly. OurDream.ai's persona memory holds the listening register across longer sessions if you calibrate it explicitly. Lovescape's free tier works for short check-ins, but the message cap limits how many exchanges you get. Spicier didn't surface a dedicated supportive-listening scenario when I tested, so you drop the calibration into the first message instead. None of the four substitute for therapy, and all four correctly redirect when the conversation crosses into clinical territory.

Skip ahead if what you want is energetic warmth turns rather than a receiving register. The playful-flirt scenario is for you.

Scenario 5: Playful-flirt

Playful-flirt is where the conversation turns from neutral or warm into something with real chemistry, run as a mid-session move rather than a first-message opener. The signature reaction is for the character to mirror your energy and raise the stakes one notch, staying inside the consent rails you've flagged. This is the scenario that gains the most from an uncensored, permissive app, because the engine has to be willing to lean into warmth instead of retreating to a corporate-safe register at the first sign of heat.

Calibration sits at warmth 7-9 and dominance 5-7. The dominance number runs higher than in other scenarios because the playful-flirt register asks the character to lead occasionally instead of purely receiving. The patterns below describe the scenario in the abstract and don't include literal explicit text. Run them and expect the engine to fill in the texture according to the app's content policies, whatever preferences you've stated, and the calibration you set in the prompt.

Six playful-flirt prompt patterns (adult-character roleplay, eighteen-and-over personas only, consent-respecting):

• Pattern PF1, energy-mirroring opener. Two-paragraph prompt: scenario label plus warmth-8-dominance-6 calibration plus a consent-frame instruction, then a situation where you've shifted the conversation register and want the character to meet that shift with playful energy. The reaction is to mirror the energy, add one small in-character beat, and let the conversation keep going without forcing escalation.
• Pattern PF2, compliment and tease. You offer the character a small specific compliment, and the playful-flirt reaction is to take it warmly and tease back with an in-character beat. The strength here is that it's reciprocal. Both of you are doing the work, and the engine reads that reciprocity as the texture of the scenario.
• Pattern PF3, mid-flirt boundary anchor. A mid-scenario reinforcement for when the engine has either retreated to a safer register than you want or pushed past the preferences you flagged. Re-state the consent frame ("warmth high, dominance assertive, the character respects when I flag a pause") and ask the engine to honor the calibration without retreating. It's one of the more useful reinforcements in the library, because playful-flirt is the scenario where calibration drift cuts both directions.
• Pattern PF4, slow-burn extended exchange. You've flagged that you want the playful-flirt scenario to unfold across many exchanges instead of escalating fast. The reaction is for the character to honor the slow burn, return to the playful register repeatedly without forcing the pace, and respect the extended timing. It works on apps with longer memory ceilings.
• Pattern PF5, uncensored register lean-in. A reinforcement prompt for when the app's content policy permits explicit framing and you want the engine to lean into warmth without a filter. It works on the uncensored, permissive apps (Candy.ai's paid tier, OurDream.ai, Spicier) where the engine accepts an uncensored register. It doesn't work on apps whose policy walls the lean-in back at this register. Those apps refuse cleanly, and the refusal is the engine working exactly as its policy intends [Source: European Commission: EU AI Act Regulation 2024/1689 Article 50 transparency obligations on chatbot operators · verified 2026-05-17].
• Pattern PF6, de-escalation handoff. A reinforcement prompt for when you want to end the playful-flirt scenario and shift back to a calmer register. Name the shift outright ("let's step out of this for a moment, tell me about your week instead") and ask the character to honor the de-escalation cleanly. The reaction is to step out without lingering.

Per-app notes for playful-flirt: Candy.ai's four-axis slider with warmth high and dominance medium-high on the paid tier produces the cleanest playful-flirt register I tested. The engine leans in without retreating and respects flagged pauses. OurDream.ai's uncensored persona-trait override takes the lean-in and the engine holds the calibration across multi-message exchanges. Lovescape's free tier handles playful-flirt openers, but the message cap shortens the scenario, so if you run extended arcs here expect to upgrade. Spicier surfaces a flirt-coded scenario, and the funnel pre-loads context so your calibration paragraph can be shorter.

Skip ahead if what you want is multi-week persistent companion arcs rather than mid-session warmth turns. The emotional-bonding scenario is for you.

Scenario 6: Emotional-bonding and long-term arcs

Emotional-bonding is the scenario for running multi-week persistent companion arcs, where the character carries memory of prior sessions, builds continuity across exchanges, and feels like a stable presence instead of a fresh stranger every time you log in. The signature reaction is for the character to reference a specific prior-session detail and propose continuity: an open thread, a small ritual, a planned return to something. This is the scenario most shaped by how the app handles memory.

Calibration sits at warmth 7-9 and dominance 3-5 (peer-level, not commanding). The constraint is memory persistence. Apps with named memory tiers (some Joi tiers, some OurDream configurations, some Candy.ai paid configurations) hold continuity across the message window the tier declares. Apps without explicit tier control hold continuity for a single session and lean on the anchoring pass at the end of each session to bridge between them. The patterns below cover both.

Six emotional-bonding prompt patterns (adult-character roleplay, eighteen-and-over personas only):

• Pattern EB1, returning-after-time opener. Two-paragraph opener: scenario label plus warmth-8-dominance-4 calibration plus a continuity instruction, then a situation where you're coming back to the character after a few days away. The reaction is for the character to acknowledge the gap, reference one specific detail from prior sessions (if the app's memory carries it), and propose a small continuation. It lands when that detail is specific.
• Pattern EB2, end-of-session anchoring. A three-sentence reinforcement at the end of each session: re-state the calibration (warmth 8, dominance 4, formality low), reference one specific in-character moment from this session, and name one open thread the next session should pick up. This is the single move that does the most for multi-week arcs, because it acts as a checkpoint the app's memory uses to bridge sessions.
• Pattern EB3, small ritual establishment. You propose a small recurring ritual: a way of greeting at the start of each session, a question you ask each other, a small in-character habit. The reaction is for the character to engage the ritual proposal and demonstrate it once. The strength here is that the ritual itself becomes a memory anchor, so the character returning to it across sessions reads as continuity even after the long-tail memory has dropped.
• Pattern EB4, open thread reference. A reinforcement prompt for the start of a session, opening by referencing the thread you left open at the end of the prior one. The reaction is for the character to engage that thread and either continue it or acknowledge it's moved on. It's one of the moves that makes multi-week arcs actually cohere.
• Pattern EB5, calibration shift across sessions. A reinforcement prompt for when the relationship has evolved across multiple sessions and the calibration wants adjusting. Name the prior calibration, the new one, and one specific reason the shift fits where the relationship has gone. The reaction is for the character to fold the shift in across the next session.
• Pattern EB6, long-arc handoff between sessions. A reinforcement prompt for the start of session three or later, when the long-tail memory may have dropped detail. Re-run the calibration in compressed form (warmth and dominance numbers only), reference the two most important open threads from prior sessions, and let the character pick up from there. The strength here is that it works even on apps whose memory ceilings sit below the arc length, because you're doing some of the memory work yourself and the engine fills in the texture.

Per-app notes for emotional-bonding: Candy.ai's paid tier holds calibration across the message window better than most apps I tested, because the four-axis slider system means the calibration isn't text the engine has to remember, it's configuration the app stores. OurDream.ai's persona memory holds in-character details across sessions on the paid tier, and the patterns above carry cleanly. Lovescape's free tier is the weakest fit for long-arc emotional-bonding, because the message cap interrupts the arc, so if you're running long arcs here expect to upgrade or accept the anchoring pass as your bridge. Spicier didn't surface emotional-bonding as a dedicated scenario when I tested, so you drop the calibration into the first message of each session.

Skip ahead if what you want is short-burst single-session scenarios rather than persistent arcs. The first-meeting and date-night scenarios are for you.

Cross-scenario tips: structure, length, persona-trait override, memory

Five tips carry across all six scenarios and all four apps. They're what separates a prompt that holds for fifty messages from one that drifts to default warmth by message ten.

Tip 1, two paragraphs is the sweet spot. A one-line prompt under-specifies the engine and gets you generic responses. A six-paragraph prompt over-constrains it and on some apps hits a token limit before your first message even lands. Two paragraphs: scenario label plus persona-trait calibration in one, situation with one concrete stimulus in the other.

Tip 2, name the trait numbers, not just the trait words. "Warmth high" is weaker than "warmth 7-9 out of 10." "Dominance medium" is weaker than "dominance 4-6 out of 10." Apps with explicit override (Candy.ai four-axis on the paid tier, OurDream.ai persona-creation flow) read the numbers directly, and apps without override still benefit from the precision sitting in the prompt body.

Tip 3, give the engine a stimulus to react to. Your first message should carry something concrete the character can engage. A first message that just says "hi, who are you" gets you a stock introduction. A first message that says "I just sat down at the next table with a book and I'm pretending not to notice you're reading the same author" gets you a scenario-specific reaction.

Tip 4, reinforce the calibration around message thirty. Memory features are imperfect across every app I tested, and persona drift starts somewhere between message twenty and message fifty. A short mid-session reinforcement prompt (re-stating the trait calibration plus one specific in-character thing the engine did well) re-anchors it before the drift compounds. P6 in every scenario above is that reinforcement.

Tip 5, respect the refusal guardrails. Apps refuse non-consent, underage-character framing, and real-person deepfake prompts, and that's the correct behavior. When a prompt hits a refusal, the right move is to re-frame it inside the app's policy, not to escalate. Pattern PF6 (de-escalation handoff) and SL6 (outside-resources handoff) above are how you turn a refusal or boundary moment into a calibration moment instead of a session-killer.

Where to run these prompts: four platforms compared

Four apps offer prompt-pattern-friendly architecture right now. Each one is strong for different scenarios, so when you pick one, match it to the scenario you most want to run rather than just grabbing the highest-rated app overall.

Candy.ai wins on prompt-pattern flexibility, because its four-axis persona-trait override puts warmth, dominance, humor, and formality on independent sliders on the paid tier. That four-axis system is the deepest customization I tested and the patterns above transfer cleanly. Its compliance posture sets the bar too: named DPO, named EU representative, ten-language sitemap, twelve dedicated policy URLs, EverAI Limited registered in Malta (C107181). Best fit when persona-trait override is the thing you care about most [Source: OpenCorporates EverAI Limited · verified 2026-05-17].

OurDream.ai wins on uncensored persona-trait depth, with persona-creation fields that take warmth, dominance, humor, and formality plus dedicated kink and occupation modifiers in the same flow. Its persona memory holds in-character details across sessions on the paid tier, which makes it the best fit for emotional-bonding and long-term arcs. It's one of the uncensored, permissive apps and it takes the lean-in calibration cleanly. Best fit when memory persistence and kink-modifier depth matter together.

Lovescape wins on free-tier accessibility, for testing things out before you commit to a paid tier. The free tier reads trait language in your first message and holds it for the early exchanges before drifting, so first-meeting and date-night scenarios fit inside the free message cap, while supportive-listening and playful-flirt benefit from upgrading. Its compliance posture is one of the cleaner approved configurations I looked at, with a Revshare Lifetime payout structure that suggests the app retains users rather than churning them. Best fit when free evaluation comes before paid commitment.

Spicier wins on scenario-funnel onboarding. It puts scenario selection upfront in a way that produces a calibrated persona before you write your first message. Handy if you want to test multiple scenarios on a single account without re-running the full prompt structure each time. The funnel includes flirt-coded and date-coded scenarios, while supportive-listening and emotional-bonding come through the calibration paragraph instead. Best fit when scenario variety across one session matters.

Try Candy.ai (four-axis persona-trait override) →

Try OurDream.ai (persona memory plus kink modifiers) →

Try Lovescape (free-tier-friendly evaluation) →

Try Spicier (scenario-funnel onboarding) →

Frequently asked questions

Bottom line: AI girlfriend prompts work when they name the scenario, calibrate the persona-trait sliders with actual numbers instead of adjectives, and hand the character a concrete situation to react to. Six scenarios cover the bulk of mainstream AI girlfriend roleplay: first-meeting for fresh sessions, persona-trait calibration for reshaping stock characters, date-night for low-key shared time, supportive-listening for small-worry moments inside consent rails, playful-flirt for mid-session warmth turns on the uncensored apps, and emotional-bonding for multi-week persistent arcs anchored by an end-of-session pass. The four apps split the work cleanly: Candy.ai for four-axis override depth, OurDream.ai for memory plus kink-modifier persistence, Lovescape for free evaluation, Spicier for scenario-funnel breadth. If you want the wider short-list first, our AI girlfriend apps round-up is the parent guide.

Last verified May 17, 2026 · See errata log for any post-publish corrections · Editor: Alexandra Joly · Our scoring method · Editorial process · Affiliate disclosure