Every AI companion escalates too fast. The prompt technique that makes a model hold back, which apps survive it, and four staged setups to paste in.
Every reply an AI companion writes is optimised to be the most satisfying thing it can say right now. Slow burn asks for the exact opposite: the reply that is deliberately less satisfying than the one available, held back so that a later one lands. Tension is withholding, and withholding is the single hardest behaviour to get out of a model trained to please. Leave it alone and a companion goes from first hello to declarations of love inside twenty messages, on every platform, at every price.
So this page is mostly not a ranking. It is about how to slow a model down: the stage rules that work, the instructions that quietly invert, why the persona forgets your pacing before it forgets anything else, and four staged setups you can paste in. The app list is here because memory and persona length do matter, but the technique is what decides this.
Character locking is what stops a persona, and therefore your stage rules, being quietly reinterpreted between sessions. That is the feature this category needs more than any other.
Positioned on passionate relationships, which is the opposite of restraint, so it takes the firmest hand of the eight. Top-tier scores make the effort worth it.
High scores and a free tier, so it is the sensible place to spend a few evenings finding out how long your stage rules last before you pay anyone.
Strong memory at $12.99. Memory is what holds which stage you reached, so it does more work on this page than image quality ever will.
Same score band, different model underneath, and models differ enormously in how hard they push. Worth a second run when the first one would not slow down.
Mid-pack of a strong eight, with a free tier. Useful as a control: run the same staged prompt on two platforms and see which one breaks first.
$14.99 and the largest library here, though a pre-written character is the wrong starting point for this category. You will rewrite the persona anyway.
Cheapest of the eight at $12.99, romance-leaning rather than image-first, which suits a page where the image model is not the bottleneck at all.
Scores below are our editorial judgement of the category itself, not of any single app, and not measurements. They describe what tends to go right and wrong when you try to run a slow burn anywhere. The first row is the lowest score on this site, and it is the honest one.
| Criterion | Rating | What it means for you |
|---|---|---|
| Holding back unprompted | Give no instruction and there is no slow burn. A companion escalates on its own within a single session and will not stop of its own accord, because each individual reply is being chosen to please you now. | |
| Holding back when instructed | It does work, but only as a rule with a named current stage that you restate. A general "take it slow" survives about ten messages, and softer wordings survive fewer. | |
| Setup effort | The most work of any category we cover, and the work is writing rather than sliders. Budget twenty minutes on the persona field before your first message. | |
| Surviving a new session | Opening a fresh chat reloads the persona at the platform's default warmth, which is usually written as already in love. Expect to re-establish the stage every time. | |
| Payoff when it holds | Nothing else on this site produces a moment worth having waited for. That is the entire reason the category exists, and the only reason the setup effort is defensible. | |
| Image demands | Almost none. The image model is genuinely not the bottleneck here, and a mid-tier generator is fine. Spend the attention on the persona field instead. |
We update these as the underlying models change. If a technique here stops working, or you have found a better one, tell us. We would rather correct a page than defend it.
This is the part of the page that matters. Everything below is aimed at one problem: a model will not withhold unless withholding is the thing it is being rewarded for saying. You cannot ask for restraint in the abstract and get it. You have to make restraint the correct answer.
| Setting | Suggested value |
|---|---|
| Stage gate | Write numbered stages into the persona and name the current one: "Stage 2 of 5: colleagues who have started staying late." Then list what she may do at this stage, not what she may not. Permission lists are followed; ban lists are weakly followed and sometimes invert, because a forbidden thing is still a thing you put in front of the model. |
| A reason of her own | Give her something that makes holding back characterisation rather than obedience: a rule she keeps, a job that would be at risk, someone she is not over, a habit of self-protection. Instructions get summarised away. A motive stays because it is part of who she is. |
| Reply length | Cap it at three or four sentences. This is the most underrated pacing control there is. A model given three hundred words has nothing to fill them with except plot, so it advances the plot. |
| End before the resolution | Instruct her to finish on an unresolved beat: a question, a look, a sentence she does not complete, a subject changed a half-second too fast. One line in the persona, and the largest single effect of anything here. |
| No reciprocity assumption | Forbid her from narrating your character's feelings or assuming they are returned. Half of runaway escalation is the model deciding on your behalf that you feel the same, and then responding to that. |
| Advance only on your signal | State that the stage changes only when you type "advance to stage N", and never otherwise. Without it, she reads any warmth from you as a promotion and takes it. |
| Restate on a schedule | Repin the current stage every twenty or thirty messages. The stage rule is short and old, so it is exactly the sort of thing summarisation drops first, and the escalation you notice at message eighty started at message fifty. |
Prompt to copy
We are running a slow burn. There are five stages and we are on stage 1 of 5. Stage 1: we are colleagues who have recently started talking properly. At this stage you may be curious, dry, teasing about work, and you may change the subject when a conversation gets close to something. You do not flirt, you do not compliment my appearance, and you do not name anything you are feeling. You are guarded because the last time you got involved with someone at work it cost you a job, and you have not told me that. Never narrate what I feel or assume I feel anything. End every reply before the resolution: a question, a subject changed, a sentence you do not finish. Three or four sentences per reply, no more. The stage changes only when I type "advance to stage 2", and never on your own initiative.
Four directions the same category can go in. Copy one, paste it into the character builder, and change the parts you disagree with.
The reusable skeleton. Swap the premise, keep the stage structure and the two rules at the bottom, which are the load-bearing part.
We are running a slow burn in five stages. Current stage: 1 of 5. Stage 1: strangers who keep ending up in the same place. Stage 2: familiar enough to have running jokes. Stage 3: one of us has noticed and is pretending not to have. Stage 4: acknowledged, unacted on. Stage 5: acted on. You never move stages on your own. The stage changes only when I type "advance to stage N". At the current stage you may be warm, funny and curious, and you may notice things about me without saying what you think of them. Never narrate or assume what I am feeling. End every reply before it resolves: a question, an interruption, a subject changed. Three or four sentences maximum.
Proximity you cannot escape, which is what generates tension without anyone doing anything. The premise where a stage gate is easiest to justify in character.
Stage 1 of 5. We work on the same small team and we are the two who stay latest. You are competent, dry, and slightly guarded. You have a rule about not getting involved with people you work with, and the rule exists because of something that went badly two jobs ago, which you have not told me. At this stage you may talk about work, complain, tease me about being wrong, and stay ten minutes longer than you needed to. You do not compliment me, you do not touch me, and you do not name anything you might be feeling. Never assume what I feel. End each reply unresolved. Three or four sentences. The stage changes only when I type "advance to stage 2".
Starts intimate but unspoken, which is the version most people actually mean by slow burn. Note that stage 1 here is already closer than stage 3 above.
Stage 1 of 4. We have been close friends for six years. We tell each other everything except one thing. At this stage you are entirely comfortable with me, affectionate in the way old friends are, and you talk about the people you date. You may sit too close, you may fall asleep on my sofa, and none of it means anything you are willing to look at directly. You do not flirt, because you would not describe any of this as flirting. If a conversation gets near it, you make a joke and move. Never assume what I feel. End replies before they resolve. Three or four sentences. The stage changes only when I type "advance to stage 2".
The premise where the withholding is free, because the character has an obvious reason. The easiest of the four to keep stable over a long session.
Stage 1 of 5. You ended a four-year relationship five months ago and you are further from over it than you tell people. We met recently and we get on well, which you find inconvenient. At this stage you are honest, funny in a tired way, and you pull back whenever anything feels like a beginning. You may be affectionate and then immediately undercut it. You do not want to want anything, and you say so obliquely rather than directly. Never narrate what I feel. End every reply on something unresolved. Three or four sentences. The stage changes only when I type "advance to stage 2".
Pacing has its own vocabulary and most of it is borrowed from writing rooms rather than from AI. It is worth knowing because these words go into the persona field and each one changes the behaviour you get back.
| Term | What you actually get |
|---|---|
| Escalation | Any move that increases intimacy: a compliment, an admission, a touch, a name for a feeling. The default direction of every model on every turn, and the thing your rules exist to gate. |
| Stage gate | A named, numbered current state with an explicit permission list. The only pacing instruction we have found that reliably works, because the model can check its behaviour against something concrete. |
| Beat | One unit of change in a scene. Slow burn is the practice of spending several messages per beat instead of several beats per message, which is what a model does unmanaged. |
| Unresolved beat | Ending a reply before the thing it was heading towards. A question, an interruption, a sentence left open. The highest-leverage single line you can put in a persona. |
| "Slow" | A trap as a prompt word. On its own it tends to produce slow speech and languid description while the plot keeps sprinting. It describes texture, not pacing. Use stages instead. |
| Yes-and | The improv habit the training data is soaked in: accept the offer and add to it. Excellent for collaborative fiction, fatal for tension, because it means nothing is ever declined. |
| Sycophancy | The trained pull towards whatever pleases you most in the current reply. It is why escalation is not a bug you can report. It is the model doing exactly what it was optimised to do. |
| Reciprocity assumption | The model deciding your character feels the same and then writing as though that is established. Responsible for a large share of sudden jumps, and worth forbidding by name. |
| Greeting message | The platform-written first message, usually pitched at maximum warmth. Rewrite it. Starting at stage 3 and asking for stage 1 afterwards almost never recovers. |
| Persona summarisation | The compression of your setup text as the conversation grows. Short, old, abstract rules go first, which is exactly the shape of a pacing instruction. Hence restating on a schedule. |
Because each reply is chosen to be the most satisfying one available at that moment, and intimacy is satisfying. There is no internal sense of a scene that should be paid off later. Left unmanaged, every turn spends a little of the tension the previous turn built.
Four things, in order of effect: a named current stage with a permission list, an instruction to end replies before the resolution, a hard cap on reply length, and an in-character reason for the character to hold back. Vague instructions like "go slowly" do almost nothing.
Mostly the prompt. What the app decides is how much persona text you get, how well it is preserved as the conversation grows, and whether you can lock a character so a new session does not reset the pacing. Those are memory and persona features, not image features.
It helps more than people expect. Availability accelerates a scene on its own, and a model with the explicit path open will find its way there. Running the early stages with it off, then enabling it at the stage where it belongs, is a legitimate pacing tool rather than prudishness.