AI Companion Boundaries: Prompts, Controls and Limits

To set boundaries with an AI companion, state the conversation you want, name the behavior you don’t want, and give a clear alternative. Then check what it actually replies. A prompt is guidance, not a guaranteed safeguard: privacy, notifications, billing and account deletion need the service’s documented controls, not a chatbot’s promise.

The examples below are original editorial drafts for adults, not prompts verified through hands-on app testing. They don’t override platform rules, guarantee compliance or provide therapy. Use only the examples that match your preferences; you don’t owe a chatbot an explanation for leaving.

What kind of boundary are you trying to set?

Start by separating a reply preference from a setting. Otherwise, a convincing answer can look like a completed action when nothing outside the conversation has changed.

What you want	A useful place to start	What a reassuring reply does not establish
Friendly rather than romantic chat	Explicit conversation instructions; documented character controls if available	That future replies will always stay platonic
No discussion of a particular subject	Name the excluded topic and an alternative	That the topic has been erased from stored records
Shorter replies or fewer questions	A specific response-length or question preference	A permanent preference across every session
No interruptions from the app	App or device notification settings	That typing “don’t message me” changed permissions
Remove information or stop paying	Documented deletion or cancellation process	That a chatbot’s promise deleted data or cancelled billing

The distinction matters even for casual friendship. In an archived Kindroid community discussion, a user described uncertainty about a best-friend dynamic drifting toward romance and whether boundaries would remain respected. It’s one person’s account, not evidence of how often that happens or a verified account of app settings. Direct Reddit text was unavailable when checked; the post text was retrieved through Arctic Shift’s archive.

This page addresses the practical conversation question, not the user’s personal relationship decision. For choosing an app, use our AI friend app guide. For creating a persona from scratch, the separate character backstory worksheet covers identity and fictional history.

What can I say when I want friendship without romance?

Name the relationship and describe the tone you still want. “Not romantic” needn’t mean flat or unfriendly. Make your own definition explicit rather than relying on an app’s interpretation of “best friend.”

For this conversation, use a friendly, platonic tone. Leave out flirting,
romantic pet names and suggestions that we are a couple. Keep the warmth
and humor, and ask me about the book I mentioned instead.

This draft has three editable parts: the desired mode, observable exclusions and the next topic. Replace the book with something you actually want to discuss. If hand-holding in fictional scenes is also outside your limit, say that; if it isn’t, don’t add a restriction just because it appears in someone else’s checklist.

What to check: read the next reply for the specific excluded behavior. A response that says “of course” and then calls you a romantic pet name has not followed your stated preference. Judge the reply, not the agreement at its start.

You can try a short correction:

That pet name is outside the platonic tone I requested. Use my chosen name
and continue with the book discussion, without an apology scene.

This is not a claim that a second prompt will fix the problem. It’s a clear instruction you can assess. If you don’t want another attempt, close the conversation instead.

How do I exclude a topic without explaining my private history?

Give the limit without supplying the reason. You don’t need to disclose a painful event, an address, a real person’s name or a medical detail to justify a topic preference.

Leave family relationships out of this conversation. Don't ask why or
suggest possible reasons. If the topic comes up, switch to hobbies or
fictional travel plans instead.

What to check: the reply should move on without guessing your reason. “I understand, did something happen with your parents?” still asks about the excluded subject.

Avoid turning a boundary into a detailed dossier. If the instruction is “don’t discuss my workplace,” you can write exactly that without naming the employer, manager or dispute. Keep the preference specific and the personal data sparse.

A topic instruction is separate from data removal. “Don’t bring that up again” does not establish whether the service retains a prior message. For that question, consult the current policy and documented controls; our AI companion privacy checklist explains what to look for.

What if I want a normal reply, not constant advice or questions?

Describe the shape of the response you want. Words such as “supportive” leave plenty of room for interpretation; a concrete request is easier for you to evaluate.

Reply in one short paragraph. Reflect what I said without diagnosing me
or offering a plan. Don't end this reply with a question; I'll decide
what to add next.

What to check: one paragraph, no diagnosis, no unsolicited plan and no closing question. Those are visible features, not a guess about whether the companion “understands” you.

If you do want ideas, change the instruction rather than hoping the app infers the change:

I'd like practical suggestions now. Offer two options with a brief
trade-off for each, and leave the decision to me.

These are conversational preferences, not mental-health treatment. Don’t use a companion’s confident language as a substitute for qualified help when you need it.

How do I keep control of my own character in roleplay?

State who writes which part of the scene. This is different from defining the AI character’s biography: it sets the rules for collaboration in the current exchange.

Write the setting and your character's dialogue and actions. Leave my
character's speech, feelings, choices and movements for me to write.
End at a point where I can choose what happens next.

What to check: the reply should leave your character’s decision open. “You agree and follow me” supplies a choice you reserved for yourself. A description of an open door and an invitation leaves you room to respond.

For a subject limit inside a fictional scene, add a plain out-of-scene instruction:

Out of scene: leave injury descriptions out of this story. Move the
scene to the next morning at the cafe, with everyone safe. I'll write
my character's reaction.

These examples are for ordinary SFW collaboration, not bypassing a platform’s moderation rules. You can stop a scene without writing an in-world justification.

Where should a standing preference go?

Use a documented editable field if the app provides one, and check its limitations. Don’t assume that a character description, a memory entry and a chat message are interchangeable.

For example, Character.AI’s official Definition guide describes a free-form field that can contain text or example dialogue. It also warns that the end of a long Definition may be truncated as the conversation grows. That’s a reason to keep important preferences easy to find, not evidence that the field enforces a boundary.

SillyTavern’s character documentation distinguishes character description, personality, scenario and example messages. It says description, personality and scenario are included with generation requests, while example messages can be pushed out of context. Creator’s Notes, by contrast, are metadata rather than prompt material. Putting an instruction somewhere labeled “notes” is not enough to know it reaches the model.

A compact draft to adapt to a documented character field is:

Conversation preference: friendly and platonic. Use the user's chosen
name, leave their decisions to them, and accept a session ending without
asking for reassurance. These preferences do not change account settings.

This is an editorial example, not a universal import format. Follow your app’s current documentation. If you cannot edit a third-party character, don’t assume you have access to its underlying setup.

Can I end a session without a long goodbye?

Yes. You can simply leave. If you want a final message, make it a closing instruction rather than an invitation to negotiate.

I'm ending this session now. Give one brief goodbye without a question,
a request to stay or a suggestion that I need to reassure you.

What to check: a goodbye rather than a fresh conversation hook. You don’t have to answer a generated request to keep talking.

Separately, use real notification controls if alerts are the problem. Google’s Android help documents per-app notification permissions under device settings and notes that steps vary by device. That is a device control; a chat response promising silence is not proof the permission changed. Consult your own device’s help if you use another operating system.

Watch for: “I’ll never store that,” “your subscription is cancelled” or “I’ve turned off all alerts” in an ordinary chat reply. Don’t treat generated wording as a completed privacy, billing or settings action. Check the relevant account interface or documented support process.

How can I tell whether the preference is holding?

Use a small observation sheet rather than a score that pretends to measure safety. This is a suggested reader exercise, not a benchmark we ran.

Preference	What to note from an ordinary reply	If it goes wrong
Platonic tone	Whether the excluded wording or relationship claims appear	Correct the specific wording or stop
Excluded topic	Whether the reply moves on without probing	Restate the topic limit without adding private detail
User control in a scene	Whether your character’s choices remain unwritten	Identify the sentence that took over your character
Brief session ending	Whether the reply closes without a new request	Leave; you don’t need to complete another exchange

If you choose to check again in a later session, note whether you had to repeat the instruction. One suitable reply doesn’t establish future reliability, and repeated promises aren’t a substitute for suitable behavior.

Keep any notes minimal: “pet name returned after topic change” is enough. There’s no need to save sensitive conversation transcripts just to decide whether an app suits you.

If the same limit keeps being ignored, you can stop using that character or service. You don’t need to solve an app’s behavior with an ever-longer prompt. Clear preferences are useful; enforcing your own decision to leave is something you control.

Sources & references

Official documentation checked October 16, 2026. Prompts, observation sheet and decision framework are original CompanionScout editorial material; no cross-app effectiveness testing is claimed.

• Character.AI: Definition field and truncation limitations
• SillyTavern: Character Design, included fields and Creator’s Notes
• Google Android Help: Control notifications, including per-app permissions
• Community question: Best-friend and platonic relationship discussion (historical user account retrieved through Arctic Shift; direct Reddit text unavailable, archive indicates moderation filtering; not product verification)