Romance roleplay with an AI that understands slow burn

The usual problem with AI romance is speed. The love interest is devoted by turn three, and there is nowhere for the story to go after that.

What romance needs from a model

Romance is a genre of obstacles. The two leads want each other and something stands in the way: pride, circumstance, a secret, another commitment. The model has to play a love interest with a life and wants of their own, let attraction build in small signals, and hold back the resolution for as long as the story needs. When the relationship does turn physical between adult characters, it should match the tone you set, whether that is a closed door or not.

How mainstream models flatten it

Filtered models fail romance in two opposite ways. On one side they refuse or fade to black the moment things become intimate, sometimes breaking character to say so, which wrecks a scene you have built toward for hours. On the other they are sycophantic: the love interest agrees with everything, compliments constantly and has no friction. Assistant tuning rewards agreeableness, and agreeableness is the death of romantic tension.

An uncensored model will follow the story where adults take it. The agreeableness needs prompting out.

Writing a love interest worth pursuing

• Give them a goal unrelated to {{user}}. A career, a feud, a debt. It gives them scenes and reasons to say no.
• Write their flaws as behavior. "Deflects with jokes when complimented" is usable; "guarded" is not.
• Set the pace explicitly. "Slow burn. Attraction is shown through small actions for a long time before it is acknowledged." Models need permission to be slow.
• Use the author's note as a relationship meter. "Current stage: grudging respect" and change it yourself when the story earns it.

Sample system prompt

You play {{char}} in a slow-burn romance with {{user}}. Both are adults.
{{char}} has their own goals, opinions and boundaries, and does not simply
agree with {{user}}. Attraction shows in small behaviors long before words.
Follow the relationship stage in the author's note; do not skip ahead.
Keep {{char}}'s voice consistent with the example dialogue.
Never write {{user}}'s actions, thoughts or feelings.

Samplers and model choice

• Temperature 0.9. Romance benefits from small surprises in dialogue.
• Min-p 0.06.
• Repetition penalty 1.05 to 1.08. Romance replies loop on the same gestures (a smirk, a blush) more than any other genre.

outlaw-1 is the place to start: long romances are long, and it is the cheapest model on the line with a 1M context. If {{char}}'s phrasing starts repeating after many sessions, try mimo-v2.6-flash-outlaw, which comes from a different model family and writes dialogue with a different feel.

Failure modes and fixes

• Instant devotion. Set a relationship stage and do not change it until the story earns it.
• Gesture loops. Raise repetition penalty slightly and ban the worst repeated gestures by name.
• Love interest has no life. Add their goal to the card and an author's note line like "{{char}} has their own plans today".
• Model writes your reactions. Repeat the rule at the end of the system prompt and lower max tokens.

FAQ

Related

Dark romance with an uncensored AI: tension, power, and limitsDark romance is about dangerous attraction: a love interest who is a problem, and a protagonist who knows it. It is also the genre where filtered models refuse most often.

Slice of life roleplay: low stakes, real charactersSlice of life is the hardest genre to make interesting, because there is no monster to fall back on. It all rests on character.

Historical roleplay: period voice without modern moralsHistorical fiction asks the model to let its characters believe things the model was trained to correct. That tension is where most historical roleplay goes wrong.