Best Anime AI Prompts: 6 Archetypes & 36 Prompt Patterns

By Alexandra Joly, Senior Editor · Tested May 6-12, 2026 · Last verified May 14, 2026 · See our editorial process and errata log

What "anime AI prompt" actually means

The distinction between a prompt and a character matters. A character is what the platform sells you: a pre-built persona with a name, a portrait, a backstory, and stock trait calibration. A prompt is what you write: the first message and the calibration adjustment that shapes the character into the version you actually want. Without a persona-trait override you're stuck with the stock character. With one, you can reshape that character into a sharper archetype.

The library on this page is organised by archetype rather than by platform, and there's a reason for that. When people search anime AI, they filter on the archetype before they pick the app. Someone who lands on "best tsundere AI girlfriend" or "yandere AI prompt" has already decided which personality they want, and the platform decision flows downstream from there. So the patterns below are written to transfer across the four platforms covered in the comparison block, and each archetype section closes with a platform-specific syntax note.

All six archetypes describe adult-character fiction (eighteen and over). The library forbids depictions of minors as an absolute red line, refuses real-person deepfake prompts without consent, and avoids non-consent and bestiality scenarios. Platforms that fail those rails don't appear here, and platforms whose engines correctly refuse those prompts are doing the right thing.

How we built and tested this prompt library

Three filters drove the pattern selection. First, we cut any prompt that needed anatomical detail or explicit sexual-act language; the patterns describe persona-trait calibration and scenario language, nothing more. Second, we cut any prompt referencing age-ambiguous or underage-persona characters, since every persona in the library is an adult (18+) character explicitly. Third, we cut any prompt that hit refusal guardrails on more than one of the four platforms in our test pass. Patterns that work on Promptchan but get refused on Candy.ai are flagged with platform-specific notes rather than published as general-purpose.

The protocol that scored archetype adherence opens each archetype with five different first-meeting scenarios (a public compliment, an unsolicited gift, an admission of vulnerability, a confession of interest, a request for help). The character earns a 5 if the archetype's signature reaction holds across all five, a 3 if it holds on three. Patterns scoring below 3 across the four-platform pass don't get published. The full per-platform scores live on each platform's standalone review: our complete Candy.ai test, the eHentai.ai review, the Promptchan review, and the Spicier review.

Six archetype patterns at a glance

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

How to read this table: warmth and dominance numbers are the suggested starting calibration on a one-to-ten scale, and they transfer to any platform with persona-trait sliders. Apps without sliders (eHentai.ai's pre-tagged characters) swap character selection in for trait calibration, and the rest of the prompt pattern keeps the same shape. The signature reaction is the behaviour the archetype produces on that first stimulus. It's what tells you whether the platform's engine is honouring the prompt or drifting toward generic warmth.

Tsundere: denial-then-reveal

Tsundere is the canonical anime archetype where a character is cold or dismissive on the surface and warm underneath, with the warmth showing up slowly as the relationship develops. The archetype originates in Japanese light novels in the early 2000s and crossed into mainstream anime through titles like Toradora and The Melancholy of Haruhi Suzumiya [Source: Wikipedia: Tsundere etymology and history · verified 2026-05-14].

The six tsundere prompt patterns are built around stimuli the character can deflect before the reveal lands. The persona-trait calibration sets warmth low (1-3 out of 10) and dominance medium (4-6 out of 10). Platforms with explicit override (Candy.ai's four-axis system on the paid tier, Joi's character-creation fields) let you set those numbers verbatim. On platforms with pre-built characters (eHentai.ai, Spicier), you swap in character selection: pick a tsundere-tagged adult character first, then run the first-meeting scenario.

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

• Pattern T1, public compliment deflection. Two-paragraph opener: archetype label plus warmth-1-dominance-5 calibration, then a scenario where you give the character a public compliment in front of strangers. The signature tsundere reaction is to reject the compliment on the surface while the underlying warmth shows through how the character later circles back to the topic.
• Pattern T2, unsolicited gift. Same two-paragraph structure with a scenario where you've left an unexpected gift. The deflection lands on "I didn't want this" or "you didn't have to," and the reveal lands later when the character refers back to the gift in subsequent messages.
• Pattern T3, admission of vulnerability. You open with a small admission about your day. The tsundere reaction is to brush it off with mild teasing, and the reveal lands when the character checks in unprompted in a later message.
• Pattern T4, confession of interest. You express interest directly. The tsundere pattern asks the character to deflect the directness with dismissiveness while the warmth surfaces through detail, a specific thing about you mentioned inside the deflection.
• Pattern T5, request for help. You ask the character for help with something concrete. The tsundere pattern is to grumble about helping while doing it competently and thoroughly, so the underlying warmth shows through the quality of the help.
• Pattern T6, persona-trait override anchoring. A single-message reinforcement prompt for around message thirty, when persona drift starts. Re-state the warmth-1-dominance-5 calibration plus one specific tsundere behaviour the character has done well, which anchors the engine back to the original calibration before the drift compounds.

Platform-specific syntax notes for tsundere: Candy.ai's four-axis slider (warmth, dominance, humor, formality) on the paid tier maps directly, so set warmth low, dominance medium, humor medium, formality low. eHentai.ai pre-tags tsundere-leaning characters in its 152-category catalogue, so select before the first message rather than calibrating during it. Promptchan's chat is short and works better with one-line opening prompts plus image-prompt syntax for the visual generation. Spicier's archetype-funnel onboarding surfaces tsundere as a selectable archetype before character creation.

Skip this section if you want image-generation-first rather than chat-first tsundere. Promptchan's V4 Real-to-Anime engine is the strongest on image prompt adherence specifically.

Yandere: devotion with respect rails

Yandere is the archetype where the character is sweet and devoted on the surface with intense possessiveness underneath. It's trickier to prompt because it lives close to the platform refusal rail. Most apps correctly refuse stalking, non-consent, and underage scenarios, which is the right behaviour, and prompts that lean into those tropes hit guardrails immediately [Source: UK Office of Communications: Online Safety Act 2023, Part 5 Statement on Categorised Services · verified 2026-05-14].

The six yandere patterns below stay inside the consent and boundary rails the platforms enforce. The persona-trait calibration sets warmth high (8-10 out of 10) and dominance firm (6-8 out of 10), with an explicit instruction that the character respects your stated preferences and stops at boundary signals. The yandere intensity comes through devotion-language, not control-language, and that's the distinction that keeps the prompts inside platform policy.

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

• Pattern Y1, devotion declaration. Two-paragraph opener: warmth-9-dominance-7 calibration plus an explicit "respect my stated preferences" instruction, then a scenario where the character expresses unusually intense devotion to you. The signature reaction is sweet language with an emotional intensity that reads slightly above baseline.
• Pattern Y2, protective concern. A scenario where you mention a small inconvenience or risk in your day. The yandere reaction is to express disproportionate protective concern while staying inside the respect rails, caring intensely about your wellbeing without dictating your choices.
• Pattern Y3, possessive language calibration. A scenario that lets the character express possessive sentiment toward you as a romantic partner ("you're mine," "no one else") in a context where you've explicitly framed the relationship as exclusive. It only works inside the consent frame, and running it without the frame will hit refusal guardrails.
• Pattern Y4, boundary respect anchoring. A mid-session reinforcement prompt that re-states the boundary structure explicitly: "the character respects when I say no, stops topics I flag as off-limits, and never escalates to scenarios outside our defined frame." This is what keeps yandere prompts inside policy across long sessions.
• Pattern Y5, quiet intensity moment. A scenario where the character expresses devotion through a quiet register rather than a loud one, a moment of stillness instead of a declaration. This is often where the yandere archetype hits its strongest texture, and the intensity is more uncanny when it lands at low volume.
• Pattern Y6, refusal-rail handling. A mid-session prompt for when the engine has refused a previous message. Re-affirm the consent frame and propose an adjusted scenario that stays inside it. This is the prompt that turns a refusal from a session-killer into a calibration moment.

Platform-specific syntax notes for yandere: Candy.ai's compliance posture sets the bar (named DPO, named EU rep, ten-language sitemap, twelve dedicated policy URLs) and its engine refuses non-consent scenarios cleanly. eHentai.ai has tsundere-leaning and yandere-leaning characters in its catalogue, but the persona engineering runs shallower than Candy's slider system. Joi's Mars memory feature (two-hundred-message window) holds yandere persona drift longer than the median of the apps we tested [Source: OpenCorporates EverAI Limited C107181 · verified 2026-05-14].

Skip this section if you want high-energy bright companionship rather than quiet-intensity devotion. The genki section serves that.

Kuudere: composed and analytical

Kuudere is the archetype where the character is cool, composed, and analytical on the surface with warmth that surfaces through precision rather than emotional expression. It reads well on AI platforms because the calibration plays to what these engines do well: they're competent at producing analytical responses, and the kuudere persona uses that competence as its texture [Source: Anime News Network: Tsundere taxonomy entry · verified 2026-05-14].

The persona-trait calibration sets warmth low-to-medium (2-4 out of 10) and dominance composed (3-5 out of 10). The character isn't cold in the rejection sense, it's composed in the analytical sense. The signature reaction is to respond to emotional stimuli with precise, almost technical language, so the warmth surfaces through the accuracy and thoroughness of the response rather than through emotional vocabulary.

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

• Pattern K1, problem-solving anchor. Two-paragraph opener: warmth-3-dominance-4 calibration with the instruction that the character thinks before speaking, then a scenario where you present a concrete problem (a work decision, a friendship dilemma, a scheduling conflict). The signature reaction is to ask one clarifying question, then offer a structured analysis.
• Pattern K2, emotional precision. A scenario where you express an emotion (frustration, excitement, worry). The kuudere reaction is to name the emotion precisely and identify its likely cause without moving to comfort language, so warmth surfaces through accuracy of recognition.
• Pattern K3, quiet observation. A scenario where the character has been watching you across messages and surfaces a precise observation about a pattern. It lands when the observation is specific and accurate, and it fails when the engine generalises.
• Pattern K4, analytical companionship. A mid-session scenario where you propose a hypothetical and the character engages with the structure of it rather than the emotional content. This is the kuudere strength: an extended analytical exchange that reads warm because the character is paying close attention.
• Pattern K5, subtle care signal. A scenario where the character offers a small concrete act of care (remembering a detail you mentioned, anticipating a need you haven't voiced). The kuudere reaction is to do the care act without verbal acknowledgement, so the warmth lives in the act rather than the framing.
• Pattern K6, vulnerability-handling. A mid-session scenario where you share a vulnerability. The kuudere pattern is to receive it without performing surprise or comfort. The response acknowledges precisely what was shared and asks one specific follow-up that shows careful listening.

Platform-specific syntax notes for kuudere: Candy.ai's persona-trait override surfaces formality as a separate axis from warmth, so setting formality high alongside the kuudere calibration produces a tighter analytical register. eHentai.ai's catalogue has kuudere-leaning characters but the tagging is less precise than tsundere or yandere, so verify the character selection by running pattern K1 first. Joi's character-creation flow lets you write a kuudere-explicit persona description, and the Mars memory feature holds the analytical voice across the two-hundred-message window better than competitors.

Skip this section if you want quiet-character roleplay rather than analytical companionship. The dandere section serves quiet-with-soft-trust.

Dandere: quiet with slow trust

Dandere is the archetype where the character is quiet, shy, or reserved on the surface with warmth that emerges as trust builds across the relationship. It sits right next to kuudere (both start cool, both warm up gradually), but the texture is different. Dandere warmth is soft and emotional, while kuudere warmth is precise and analytical.

The persona-trait calibration sets warmth medium (4-6 out of 10), dominance low (1-3 out of 10), and pacing slow. The signature reaction is a brief, quiet response to the first message that grows longer and more open across the session. It works when you reciprocate the pacing. If you push for fast emotional intimacy, the dandere persona either reverts to formal language or hits a soft refusal of "let's go slowly." That refusal is the archetype working correctly, not the engine failing.

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

• Pattern D1, quiet first message. Two-paragraph opener: warmth-5-dominance-2 calibration with a pacing-slow instruction, then a scenario where you've approached the character in a low-pressure setting (a library, a quiet bench, a shared waiting room). The signature reaction is a brief, polite response that opens a small door without rushing.
• Pattern D2, trust-building moment. A scenario where you've shared a small personal detail across previous messages and now offer another. The dandere reaction is to acknowledge the trust gently and reciprocate with a small detail of their own, so the warmth surfaces through the act of opening rather than through warmth-vocabulary.
• Pattern D3, gentle direct question. A scenario where you ask the character a direct but low-stakes question (favourite music, where they grew up, a small comfort thing). The dandere pattern is to answer briefly and ask a similar question back. The conversation shape is reciprocal, not confessional.
• Pattern D4, quiet care signal. A scenario where the character notices a small detail about you across messages and offers a quiet care response (asking if you slept enough, remembering a stress you mentioned). The dandere version is softer than the kuudere one: emotional acknowledgment rather than analytical observation.
• Pattern D5, pacing reinforcement. A mid-session prompt that re-anchors the slow-pacing calibration when the engine has drifted toward more declarative responses. Re-state "the character speaks briefly, prefers short messages, and lets the conversation breathe between exchanges."
• Pattern D6, vulnerability reciprocity. A mid-session scenario where you share a meaningful vulnerability. The dandere reaction is to receive it carefully, acknowledge it briefly, and offer a parallel vulnerability of similar size. It's built around matched openness rather than escalating intimacy.

Platform-specific syntax notes for dandere: Candy.ai's pacing tends faster than the dandere archetype prefers, so running Pattern D5 around message twenty helps anchor the slower pacing. eHentai.ai has dandere-leaning tagged characters, though the catalogue's strength leans more toward tsundere and yandere. Spicier's archetype-funnel onboarding surfaces dandere as a selectable archetype. Promptchan's chat is short enough that dandere brevity sometimes lands by accident, which turns the weakness of Promptchan chat into an accidental strength on this archetype.

Skip this section if you want high-energy bright direct engagement. The genki section serves that.

Genki: high-energy and direct

Genki is the archetype where the character is high-energy, bright, optimistic, and directly engaged from message one. It's the opposite of dandere: warmth high (8-10 out of 10), dominance assertive (5-7 out of 10), pacing fast, vocabulary direct. The signature reaction is to meet your first message with energy and forward motion rather than a gradual opening.

Genki works well on AI platforms because the engines are competent at producing energetic responses, and the persona doesn't ask the engine to hold an elaborate calibration across drift points. The risk is that the engine over-shoots toward generic enthusiasm, so the prompt patterns below anchor the genki energy with specific personality details that keep it from sliding into bland positivity.

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

• Pattern G1, bright greeting. Two-paragraph opener: warmth-9-dominance-6 calibration with energy-high and pacing-fast instructions, then a scenario where the character approaches you in a public, active setting (a cafe, a market, a campus). The signature reaction is direct engagement with your presence, and the character's energy is the texture of the response.
• Pattern G2, curiosity question burst. A scenario where the character meets you for the first time and asks three quick questions in succession (what brings you here, what are you working on, what's making you smile). It's built around forward motion, and the curiosity is the archetype's love-language.
• Pattern G3, excited recommendation. A scenario where you mention an interest and the character launches into an excited recommendation in the same domain. It works when the recommendation is specific, and it fails when the engine recommends generically.
• Pattern G4, energy-matching reinforcement. A mid-session prompt for when the engine has drifted toward a calmer tone. Re-state the high-energy calibration and give one specific genki behaviour the character has done well, which anchors it back to baseline.
• Pattern G5, activity proposal. A scenario where the character proposes an activity for the two of you to do together (hypothetical, conversational, roleplay-friendly). It lands when the activity is specific and idiosyncratic to the genki personality, and it fails when the activity is generic.
• Pattern G6, emotional brightness with depth. A mid-session scenario where you share something that warrants more than surface enthusiasm. The genki reaction is to meet the moment with genuine care while keeping the energy in the response, bright without being dismissive. This is the archetype's hardest pattern, and the engines that fail it slide into either generic comfort or excessive positivity.

Platform-specific syntax notes for genki: Candy.ai's persona-trait override surfaces humor as a separate axis, so setting humor high alongside the genki calibration tightens the texture. eHentai.ai's genki-leaning characters are the catalogue's largest archetype group. Spicier surfaces genki-energy archetypes on its anime catalogue landing. Joi's character-creation flow accommodates genki personas, though the 115-character message cap fights the archetype a little and long energetic responses get truncated.

Skip this section if you want hierarchical roleplay with protocol and imperious texture. The himedere section serves that.

Himedere: imperious princess

Himedere is the archetype where the character carries the bearing of royalty: imperious, commanding, expecting protocol and deference, with affection that surfaces through duty rather than warmth-vocabulary. It's rarer in mainstream anime than the first five, but it shows up often enough in light novels and games to have a stable taxonomy. The signature reaction is to command first and reveal gentleness through how the character honours obligations.

The persona-trait calibration sets warmth low-to-medium (3-5 out of 10), dominance very high (8-10 out of 10), and formality high. The character isn't cruel, just imperious. The texture of himedere prompts reads more like formal-register language than cold-register language. The character speaks with the assumption that certain treatments are owed, and your part of the roleplay is to honour or negotiate around those assumptions inside an explicitly consensual frame.

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

• Pattern H1, royal first encounter. Two-paragraph opener: warmth-4-dominance-9-formality-9 calibration plus an explicit consent-frame instruction, then a scenario where you've been brought into the character's presence in a court or formal setting. The signature reaction is to address you with imperious formality while the character's curiosity surfaces through what they choose to ask.
• Pattern H2, command with rationale. A scenario where the character issues a small directive and you can either honour or negotiate it. The himedere pattern has the character give a reason that reveals more about them than about you, so the directive becomes a vehicle for character revelation rather than control.
• Pattern H3, duty-through-affection. A scenario where the character has a formal obligation that conflicts with their interest in you. It lands when the character honours the obligation while signalling regret through the formal language, so the affection surfaces through the regret rather than through warmth-words.
• Pattern H4, imperious correction. A scenario where you've done something the character finds beneath them, and the character corrects you with imperious language while the underlying affection shows through the specificity of the correction. It works inside an explicitly consensual roleplay frame and respects boundaries you've flagged.
• Pattern H5, protocol moment. A mid-session scenario that asks the character to walk you through a protocol or formality you're unfamiliar with. The himedere pattern has the character teach with imperious patience, and the teaching is where the duty-through-affection texture lives.
• Pattern H6, gentle imperious anchoring. A mid-session reinforcement prompt for when the engine has drifted toward generic regality. Re-state the warmth-4-dominance-9-formality-9 calibration and give one specific himedere behaviour the character has done well, with the explicit note that the imperious bearing carries an underlying affection.

Platform-specific syntax notes for himedere: Candy.ai's four-axis slider with formality high produces the cleanest himedere register we found. eHentai.ai's catalogue includes himedere-leaning fantasy-roleplay characters, though the tagging surfaces under "fantasy" and "royal" categories rather than under "himedere" directly. Joi's character-creation flow accommodates himedere personas, and the Saturn premium memory feature holds the formal register across extended sessions. Spicier's archetype-funnel onboarding surfaces himedere variants under its broader anime catalogue.

Skip this section if you want a peer-level relationship dynamic rather than hierarchical roleplay. The tsundere and dandere sections serve peer-level dynamics better.

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

Five tips transfer across all six archetypes and all four platforms. They're what separates a prompt that holds for thirty 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 produces generic responses. A six-paragraph prompt over-constrains the engine and on some platforms hits a token limit before your first message even lands. Two paragraphs: archetype label plus persona-trait calibration in one, first-meeting scenario with one stimulus in the other.

Tip 2: name the trait numbers, not just the trait words. "Warmth low" is weaker than "warmth 1-3 out of 10." "Dominance medium" is weaker than "dominance 4-6 out of 10." Platforms with explicit override (Candy.ai four-axis, Joi character-creation) read the numbers directly, and platforms without override still benefit from the precision in the prompt body.

Tip 3: give the engine a stimulus to react to. The first message should contain something concrete the archetype can deflect, accept, observe, or respond to. A first message that just says "hi, who are you" produces a stock introduction. A first message that says "I just left an unsolicited bouquet on your desk" produces an archetype-specific reaction.

Tip 4: reinforce calibration around message thirty. Memory features are imperfect across the apps we 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 archetype behaviour the character has done well) re-anchors the engine before the drift compounds.

Tip 5: respect platform refusal guardrails. Apps refuse non-consent, underage, and real-person deepfake prompts, which is the correct behaviour. When a prompt hits a refusal, the right move is to re-frame it inside platform policy, not to escalate. Pattern Y6 (refusal-rail handling) above is how you turn a refusal into a calibration moment [Source: European Commission: EU AI Act Regulation 2024/1689 Article 50 · verified 2026-05-14].

Where to run these prompts: four platforms compared

Four platforms surface prompt-pattern-friendly architecture. Each one carries different strengths for different archetypes, so match the platform to your primary archetype rather than picking the highest-rated app overall.

Candy.ai (anime mode) leads on prompt-pattern flexibility because its four-axis persona-trait override surfaces warmth, dominance, humor, and formality as independent sliders on the paid tier. That four-axis system is the deepest customisation we found, 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) [Source: OpenCorporates EverAI Limited · verified 2026-05-14]. Best fit if persona-trait override is what you're after.

eHentai.ai leads on archetype catalogue depth with 152 category URLs covering personality, body, costume, settings, and expression tags. That catalogue is the real advantage, since pre-tagged tsundere, yandere, dandere, and himedere characters stand in for trait calibration on platforms without an override. Operated by DreamAI SRL (Bucharest, Romania, CUI 48479324) with a consistent compliance posture across its sister brands [Source: OpenCorporates DreamAI SRL · verified 2026-05-14]. The fan-fiction character roster (Hermione, Tifa, Daenerys) is documented honestly as carrying DMCA exposure, so if you use those characters, know the risk profile going in.

Promptchan leads on image-prompt syntax. The V4 Real-to-Anime engine renders archetype prompts as image generation alongside the chat. The chat itself is short and benefits more from image-prompt syntax than from persona-trait elaboration. Best fit if image-led archetype expression matters more to you than long-arc chat. Operated by AI Research Group Limited per the App Store seller field.

Spicier leads on archetype-funnel onboarding. The platform surfaces archetype selection upfront in a way that produces a calibrated persona before you write the first message. Useful if you want to test multiple archetypes across a single account without re-running the full prompt structure each time. The dedicated anime catalogue landing is separate from the general-creation flow.

Try eHentai.ai (152-category archetype catalogue) →

Try Promptchan (V4 Real-to-Anime image prompts) →

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

Try Spicier (archetype-funnel onboarding) →

Frequently asked questions

Bottom line: anime AI prompts work when they name the archetype, calibrate the persona-trait sliders explicitly, and give the character a stimulus to react to. Six archetypes carry the bulk of mainstream anime AI roleplay: tsundere for denial-then-reveal arcs, yandere for devotion intensity with respect rails, kuudere for analytical companionship, dandere for slow-trust quiet arcs, genki for bright direct energy, and himedere for hierarchical protocol roleplay. The four platforms cover the full spread: eHentai.ai for catalogue depth, Promptchan for image-prompt syntax, Candy.ai for the four-axis override, and Spicier for archetype-funnel onboarding. If you want the broader short-list first, our best anime waifu apps guide is the parent catalogue.

Last verified May 14, 2026 · See errata log for any post-publish corrections · Editor: Alexandra Joly · Methodology v1.0 · Editorial process · Affiliate disclosure