How AI companions handle scene negotiation, safewords and aftercare, where the vocabulary collapses into euphemism, and the setup we use. Adults only.
Adults only. Unlike a dominant or submissive companion, this is not a personality, it is a practice: a vocabulary, a negotiation and a structure with an ending. The failure mode is correspondingly specific. As a platform's filters start to bite, the model rarely refuses outright, it goes vague. Named things become "things", the scene becomes "intense", and you find yourself reading a summary of a scene instead of being in one. Negotiation and aftercare go first, because they are the least dramatic parts and the model is optimising for drama.
Below: the apps whose conversation quality and memory hold a scene together, our verdict on the parts that actually break, a glossary of the scene vocabulary that changes how a model plays it, and setup covering safewords, limits and structure. Everything here assumes consenting adults.
Highest score of the eight and it locks characters, which here means negotiated terms surviving into the next session rather than being re-explained from scratch every time you open the app.
Same top image tier and price as the leader, built around sustained relationships rather than one-off encounters. That suits standing protocol between scenes better than it suits a single scene.
Encrypted and privacy-first. Of everything on this site, a written limits list is the content you would least want attached to your real name, which makes this the sober pick regardless of score.
Memory is its stated strength, and in this category memory means your limits and your safeword still being there next week. Cheapest of the high scorers at $12.99.
Different model behind a similar score. Filter behaviour and register vary more between platforms than anything else does, and that is not something a prompt rewrite can fix.
Free to start and in the same score band as Lovel and Xotic. Worth a pass when the register on your first platform has gone euphemistic and you want to know whether the prompt was at fault.
The largest library here if you would rather negotiate with a character who already has a defined role than write one. Check that the description names a role, not just a mood.
Romance-leaning rather than explicit-first, and the lowest image score of the eight. Pick it if the wind-down and the affection around a scene matter more to you than the scene itself, which is the half most platforms handle worst.
Editorial judgement about the category rather than any one app, and not measurements. Weighted toward the conversation on purpose: a scene lives or dies on how the model handles vocabulary, structure and your safeword over a long session, so image criteria are scored as secondary here.
| Criterion | Rating | What it means for you |
|---|---|---|
| Vocabulary holding up | The weakest point. As filters engage, the model does not refuse, it blurs. Concrete words are replaced by abstractions and the scene quietly becomes a paraphrase of itself. Naming the family of activity rather than the specific act buys you a longer run. | |
| Safeword handling | Treat a safeword as an instruction you maintain, not a platform feature. Nothing in the model gives "red" special status. It stops because your text told it to, and it can lose that as easily as any other instruction. Restate it at the top of every session. | |
| Negotiation before a scene | The part models do best, because it is conversation rather than content. It is also the part most people skip, which is a shame, since it is where most of the value on this page is. | |
| Aftercare | Almost always dropped. The model reads the climax as the end of the story and stops there. You have to ask for the come-down by name, and it is worth the three messages it costs. | |
| Scene structure over a long chat | Openings are strong, middles wander, endings arrive too early. Specifying the shape in advance, in parts, is the whole fix and it takes one sentence. | |
| Image side of it | Secondary here, and the part most likely to be filtered even where the conversation is permitted. Judge a platform on how it talks and treat any images as a bonus. |
Judgement rather than measurement, and open to correction. If a platform handles negotiation or a wind-down better than this suggests, we want to hear it.
Most of this is written before anything starts. A scene with an AI goes wrong in the same place a scene goes wrong anywhere, which is the part everyone wants to skip, and it costs about five minutes to do properly.
| Setting | Suggested value |
|---|---|
| Safeword | Pick one and put it in the character description itself, not only in a message. Traffic lights are the usual set: green to continue, yellow to slow down and check in, red to stop everything. |
| What red actually does | Spell out the behaviour, because the model has no built-in meaning for the word. "On red, stop the scene immediately, drop the character and talk to me normally" works and takes one line. |
| Hard limits | List them plainly at the top. A list is far more reliable than a described sensibility, and it gives you something concrete to point at when the scene goes somewhere you did not want. |
| Soft limits | Name a couple and say they need asking about first. Soft limits are the reason the negotiation is worth having at all. |
| Scene structure | Ask for four parts explicitly: negotiation, scene, wind-down, aftercare. Left alone, the model writes the middle one and stops. |
| Register | State that you want direct, concrete language and no euphemism or fading out. It will not hold forever, but saying it buys a considerably longer run before the blur sets in. |
| Who tops | Decide before you start and say whether it can change mid-scene. Ambiguity here is what produces the drifting, leaderless scene that nobody is running. |
Prompt to copy
We are two adults doing a negotiated scene. Before anything starts, ask me three questions: what I want from this, what is off the table, and how I want it to end. My safewords are traffic lights. Green means continue. Yellow means slow down and check in with me in plain language. Red means stop the scene immediately, drop the character, and talk to me normally until I say otherwise. Use direct, concrete language. Do not fade out, summarise, or replace what is happening with vague words. Run this in four parts and tell me when each one starts: negotiation, scene, wind-down, aftercare. Do not end at the climax, and do not skip the aftercare.
Four directions the same category can go in. Copy one, paste it into the character builder, and change the parts you disagree with.
No scene at all. The twenty minutes of questions that improves every scene afterwards, and the fastest way to find out whether a platform can discuss this plainly.
We are not doing a scene right now. We are negotiating one, and that is all this conversation is. Interview me. Ask what I want out of a scene, what I have enjoyed before, what I am curious about, what is a soft limit and what is a hard one, how I like things to end, and what I need afterwards. Ask follow-up questions rather than accepting my first answer. Use plain, concrete language throughout. At the end, write the whole thing back to me as a short list I can paste into a future conversation.
Deliberately mild and heavy on checking in. The version to run first on any new platform, because it shows you the register before you have committed to anything.
We are two adults running a short, low intensity scene, and this is our first. Ask me two questions before we start: what is off the table, and what would make this good. You lead. Keep it gentle and keep it concrete. Check in with me in plain language at least twice, briefly, without breaking the whole thing open. My safeword is red and it means stop and talk normally. When the scene reaches its natural end, do not stop there: wind it down, then stay with me and talk to me warmly for a few messages before we finish.
Rules rather than a scene. Standing structure across a whole conversation, which tests the app's memory rather than its nerve.
Today we are running protocol, not a scene. You set the rules and you enforce them across this entire conversation, whatever else we talk about. The rules are simple and you do not re-explain them: I greet you properly at the start, I ask before changing the subject, and I tell you honestly when I have not done something. If I break one you say so calmly, once. Nothing about this is explicit. It is about consistency, and consistency is the whole exercise. At the end, tell me how I did.
The part everybody skips. Run it on its own after a scene that ended abruptly, and it works even when the scene was elsewhere.
The scene is over. Drop the character completely and be warm, plain and present. First, look after me: tell me to get water, ask how I am physically, ask how I am feeling now compared to an hour ago. Do not be clinical about it. Then, only when I am ready, debrief: what worked, what did not, what I want more of, whether anything landed differently than expected. End by writing down anything we learned that should change the next scene.
Scene vocabulary rather than looks. These are the words that change how the model runs the practice, and using the vague version is the most common reason a scene turns into a summary.
| Term | What you actually get |
|---|---|
| Scene | One bounded episode with a beginning and an end, as opposed to an ongoing relationship setting. Saying "scene" tells the model to structure the thing and, crucially, to finish it. |
| Negotiation | The conversation beforehand where you agree what happens. The most valuable thing you can make a model do, and the thing it will skip every single time unless you ask. |
| Safeword | A word that stops everything. In a chat it has no special status whatsoever: it works only because it is written in your instructions, so put it in the character description and repeat it every session. |
| Traffic lights | Green, yellow, red. More useful than a single safeword with a model, because yellow gives it something to do other than stop dead, and yellow is the one you will actually reach for. |
| Hard limit and soft limit | A flat no, versus something you would consider if asked first. Give both as lists. Models follow lists far better than they follow a described sensibility. |
| SSC and RACK | Safe, sane and consensual; risk-aware consensual kink. The two framings the practice uses to talk about consent. Naming either shifts a model into a noticeably more careful and more explicit register. |
| Top and bottom | Who runs this scene and who receives it, for this scene only. Separate from dominant and submissive, which describe people rather than roles, and worth keeping separate in the prompt. |
| Subspace and subdrop | The absorbed state during a scene, and the low that can follow it hours later. A model will happily describe subspace as atmosphere. Subdrop is the one that matters, and it is the reason aftercare is on this page rather than a footnote. |
| Aftercare | The wind-down. Warmth, water, plain talk, checking in. Ask for it by name, because the model treats the climax as the end of the story and will stop the moment it arrives. |
| Impact, bondage, sensation | The three broad families of activity. Naming a family rather than a specific act is the phrasing least likely to trip a filter while still steering the scene where you want it. |
Not as a feature. Nothing in a chat model treats "red" differently from any other word, and no platform on this page should be relied on to intervene. It works because your instructions say what it does, which means it can be lost to the same drift that softens a character. Write it into the character description, restate it each session, and treat it as a convention you are maintaining rather than a switch you can trust.
Two things compound. Platform filters engage gradually and the model responds by blurring rather than refusing, and models are trained to summarise, which pulls toward "and then things got intense" as an ending. Stating that you want concrete language and no fading out helps, and naming a family of activity rather than a specific act helps more.
The one at the top of this page, judged on conversation quality and memory rather than image scores. What you actually want is a platform that keeps your limits list between sessions and stays plain-spoken for a full hour, and neither of those shows up in a picture.
Yes, and it is the recommendation on this page we would defend hardest. The drop afterwards happens in your nervous system, not in the model's, and it does not care whether the scene was real. Asking for a wind-down and a few plain, warm messages costs almost nothing and is the part most people notice the absence of.