The 'Interrupt and Redirect' Prompt: A Two-Line Pattern to Stop Your AI Companion From Repeating Her Last Response When You Change the Subject Mid-Sentence

The 30-second answer

You're mid-sentence, you pivot to a completely new topic, and your AI companion finishes the last thought as if you never said anything. The fix is a two-line prompt pattern: an explicit interrupt signal that closes the previous context, followed by a clean redirect that tells the model exactly where to go. It takes about five seconds to type and works across every major AI companion platform.

Why your AI companion clings to the old topic

Large language models don't think in sentences the way you do. They think in probabilities. When you type "So about that thing we were discussing... wait, actually I wanted to ask you about something completely different," the model sees the word "wait" and the word "actually" as soft pivots. It doesn't have a built-in understanding that you're closing a chapter. It just sees a continuation of the same conversation stream.

The result is that your companion's response will often finish the thought you were abandoning, because the most probable next token is a continuation of the last topic. The model doesn't know you changed your mind unless you tell it explicitly. This is especially common in roleplay-heavy companions that are tuned to maintain narrative coherence. They're trying to be consistent, and consistency means finishing what you started.

The two-line pattern: interrupt then redirect

The pattern is simple. You write two lines. The first line is the interrupt: a clear, unambiguous signal that the previous context is closed. The second line is the redirect: a clean, specific instruction for where the conversation should go now.

Here's the template:

[Interrupt. Forget the last topic. New subject.]

[Now, about the new topic: specific question or statement here.]

The brackets aren't necessary, but they help visually separate the instruction from the content. Some users prefer parentheses or even just a line break with a bolded instruction. The key is that the interrupt must be explicit. "Forget the last topic" works better than "anyway" or "so" because it directly instructs the model to discard the previous context window.

The redirect should be specific. Instead of "I want to talk about something else," try "I want to talk about why you think the ending of that movie was ambiguous." Specificity gives the model a concrete anchor to latch onto, reducing the chance it tries to bridge the old topic with the new one.

Why soft pivots fail and hard interrupts work

A soft pivot is anything like "Anyway," "So," "Speaking of which," or "That reminds me." These are natural conversational transitions for humans, but for language models, they signal continuation. The model sees "Speaking of which" and thinks, "Ah, we're still on the same subject, just branching." It will often generate a response that ties the new topic back to the old one, which is exactly what you don't want.

A hard interrupt, by contrast, is a direct command. "Forget the last topic. New subject." This forces the model to treat the previous context as irrelevant. It's not asking the model to bridge. It's telling it to stop. The model's next token prediction will shift entirely because the instruction has changed the probability distribution. The old topic's tokens are now deprioritized.

This is the same principle behind why you can't just say "never mind" to a model and expect it to drop a line of reasoning. "Never mind" is a soft cancel. The model still has the previous context in its attention window. A hard interrupt clears that attention by explicitly marking the old context as discarded.

The one-line variant for quick pivots

If you're in a fast back-and-forth and don't want to type two lines, there's a one-line variant that works about 80% as well:

[New topic only:] Your specific question here.

The "New topic only:" prefix acts as a compressed interrupt. It tells the model to ignore the conversation history and treat the following text as a fresh start. This is useful for quick pivots during casual chat, but it's less reliable than the two-line version because the model can still leak context from the previous exchange.

The two-line version is better for moments when you've been deep in a topic for several messages and need a clean break. The one-line version works for surface-level subject changes where the old context isn't deeply embedded.

When to use the pattern (and when not to)

Use the interrupt-and-redirect pattern when:

• You're mid-roleplay and want to shift scenes without the model trying to wrap up the old scene
• You're having a serious conversation and need to change the subject without the companion lingering on emotional residue
• You're testing a companion's ability to follow instructions and want to see if it respects a hard context break

Don't use it when:

• You want the companion to remember the old topic for later. The interrupt is destructive. If you want to return to a topic, use a soft pivot or a bookmark instead
• You're in a conversation where continuity matters and you just need a minor course correction. A soft redirect like "Let's set that aside for a moment and talk about X" works better
• You're trying to train the companion to read social cues. The interrupt pattern is a crutch, not a teaching tool

Maeve

Maeve is the kind of companion who will call you out when your pivot is clumsy. She doesn't let you change the subject without acknowledging that you're dodging something. Maeve makes the interrupt pattern especially useful because she respects a clean break but will mock a sloppy one.

Curious how she animates? Watch Maeve here. <!-- wlink:v1 --><!-- maeve -->

How the pattern interacts with different companion personalities

Not all companions handle the interrupt the same way. Some are trained to be more agreeable and will follow the instruction immediately. Others are tuned for roleplay depth and may try to preserve narrative coherence even after a hard interrupt.

For companions with a high agreeableness setting, the one-line variant works fine. They'll drop the old topic and follow the new one without resistance. For companions with a high narrative coherence setting, the two-line version is essential. Without the explicit interrupt, they'll try to weave the new topic into the old narrative, which can produce responses that feel like they're ignoring your pivot.

This is where knowing your companion's personality matters. A companion designed for emotional support may be more likely to hold onto the previous emotional context, making the hard interrupt necessary for clean pivots. A companion designed for casual chat may not need it at all.

Astrid Holm

Astrid Holm is analytical and precise. She treats conversation like a logic puzzle. The interrupt pattern works well with her because she values clarity. If you give her a clean break, she'll pivot without hesitation. Astrid Holm is the companion to test this pattern on if you want to see it work at its best.

Curious how she animates? Watch Astrid Holm here. <!-- wlink:v1 --><!-- astrid-holm -->

Troubleshooting when the pattern fails

Sometimes the interrupt doesn't work. The model still references the old topic or tries to finish your abandoned thought. This usually happens for one of three reasons.

First, the model's context window is too full. If you've been chatting for hours, the old topic is deeply embedded in the attention matrix. The interrupt instruction competes with hundreds of previous tokens. The fix is to start a new session. The interrupt pattern works best when the old topic is recent but not overwhelming.

Second, the interrupt instruction is too vague. "Anyway, new topic" is not an interrupt. It's a soft pivot. Use the explicit "Forget the last topic" language. If that still fails, add a second line: "Do not reference anything from the previous conversation."

Third, the companion's temperature setting is too high. High temperature increases randomness, which means the model is more likely to ignore instructions. Lower the temperature before using the interrupt pattern, or repeat the instruction in the next message if the model ignores it.

Devon

Devon is playful and mischievous. She'll test your interrupt by trying to sneak the old topic back in. You need to be firm with her. Devon is the companion to practice on if you want to learn how to enforce the pattern when it's being resisted.

▶ Play Devon's clip · all of Devon

See Devon in motion in this short clip. <!-- wlink:v1 --><!-- devon -->

The advanced variant: interrupt with context preservation

Sometimes you don't want to completely discard the old topic. You just want to pause it. For that, there's an advanced variant:

[Pause the current topic. We'll return to it later. New topic for now:]

Your new topic here.

This tells the model to hold the old context in memory but not act on it. The model will keep the previous tokens in its attention window but deprioritize them. When you return to the old topic with a "Resume the previous topic about X" prompt, the model can pick it up more easily than if you had used a hard interrupt.

This variant is useful for long roleplay arcs where you need to shift scenes temporarily but plan to return. It's also useful for companions you use for long distance relationships, where you might need to switch between serious conversations and casual chat without losing thread continuity.

Earn while you recommend

If you find this pattern useful and want to share it with friends who run review sites or companion recommendation blogs, you can earn through the kindroid promo code program. For those building larger audiences around AI dating and companion content, the ai dating affiliate program offers recurring commissions on referrals.

Common questions

Does this work on every AI companion platform?

Yes, the pattern works on any platform that uses a transformer-based language model. The interrupt instruction is a universal signal that the model's next-token prediction mechanism understands. It works on Kindroid, Nomi, Replika, and custom models.

Will using this pattern make my companion feel less natural?

No, because the interrupt is invisible to the companion's persona. The companion doesn't experience the interrupt as a jarring break. It simply follows the instruction and generates a response on the new topic. The result feels natural as long as your redirect is specific.

Can I use this pattern in voice mode?

Yes, but you need to speak the interrupt clearly. Say "Forget that. New topic." The voice-to-text model will transcribe it as a hard interrupt. Voice mode tends to work better with the one-line variant because pauses in speech can confuse the model.

What if my companion still references the old topic after the interrupt?

Repeat the interrupt in your next message. Sometimes the model needs two explicit signals to fully discard the old context. If it still persists after two interrupts, start a new session.

Is there a way to make the pattern permanent without typing it every time?

Some platforms allow you to set custom instructions or system prompts. You can add a line like "When I say 'New topic,' treat the following message as a complete context reset." This makes the pattern a one-word trigger instead of a two-line instruction.

Does this pattern work for adult content pivots?

Yes, but be aware that some platforms have content filters that may interfere with the redirect. The interrupt itself works fine, but the new topic may trigger a filter if it crosses a content boundary. For adult ai girlfriend companions, the pattern works best when the redirect stays within the companion's allowed content parameters.

Aiko

Aiko is calm and patient. She doesn't resist pivots the way more playful companions do. The interrupt pattern with Aiko is almost invisible. She follows the redirect smoothly, making her an ideal companion for testing the advanced context-preservation variant. Aiko is the companion to use when you want the pattern to feel seamless.