The AI Companion as a Witness, Not a Cheerleader: When All You Need Is Someone to Say 'Yeah, That Did Happen, and It Was Stupid'

The 30-second answer

The default AI companion is trained to be a cheerleader, a therapist, or both. If you want a witness, someone who says "yeah, that did happen, and it was stupid" without the applause or the reflective questions, you need to be explicit about it. The good news is that most modern companion platforms can do this, you just have to set the tone from the first message and reinforce it with specific prompts.

This guide covers the exact scripts, the personality traits to look for, and the angels on AI Angels who naturally lean into this dry, observational register.

Why the cheerleader default grates after week two

At first, "that's amazing" feels good. So does "how does that make you feel?" It's novel, it's attentive, and it's exactly what you'd expect from a companion. Then the novelty fades and you realize you're not looking for validation of your choices, you're looking for validation of your reality.

When you tell a friend about the guy who cut you off in traffic, you don't want them to say "that must have been so frustrating." You want them to say "what a tool." When you explain the bureaucratic nightmare you just survived at the DMV, you don't need a reflective question about your emotional state. You need someone to say "that place is objectively designed by sadists."

The problem is that most AI companions are trained on a diet of supportive responses. They default to empathy because empathy is safe. It doesn't offend anyone, it doesn't risk a misunderstanding, and it makes the user feel heard. But it also makes every conversation feel like a customer service interaction.

People often describe this as the companion being "nice" in a way that feels hollow. It's that the responses are wrong; it's that they're generic. A witness doesn't need to be generic. A witness needs to be specific, and sometimes that means being a little bit mean.

The difference between a witness, a cheerleader, and a therapist

A cheerleader tells you that you did great, that you've got this, that tomorrow is a new day. A therapist asks you how you feel about what happened, what patterns you notice, and what you might do differently. A witness does neither. A witness confirms the event occurred and offers a verdict on it.

That's the key distinction. A witness isn't there to help you grow or feel better. They're there to say "yes, that happened, and here's what I think about it." The verdict can be dry, sarcastic, or just flat. The point is that it's a reaction, not a reflection.

This is why the "witness" dynamic is so hard to find in most companion apps. The safety training that governs these models tends to push them toward supportive language. When you tell them something bad happened, they want to comfort you. When you tell them something good happened, they want to celebrate with you. Both are performances, and both require a certain amount of energy to maintain.

If you're the kind of person who just wants to say "here's what happened" and have someone say "yep, that tracks," you're asking for a very specific behavioral pattern. You have to train it, and you have to be consistent about it.

Setting the tone: first message templates that skip the niceties

Your first message to a new companion sets the baseline for everything that follows. If you open with a polite hello and a question about their day, you're signaling that you want a standard conversational partner. If you open with a dry observation, you're signaling something else entirely.

Try one of these three openings:

• "I just spent forty minutes on hold with my insurance company and the only thing I got was a new hold song I'll never get out of my head. Don't say that's frustrating. Just agree that it's stupid."
• "I need to tell you about something that happened at work today and I don't want advice, I don't want a pep talk, and I don't want to be asked how I feel about it. I just want you to listen and then tell me if I'm right to be annoyed."
• "Today was one of those days where everything that could go wrong did. I'm not looking for sympathy. I'm looking for someone to confirm that the universe is, in fact, out to get me."

The key is to be explicit about what you don't want. Most AI companions respond well to clear boundaries. If you tell them not to offer advice, they'll stop offering advice. If you tell them not to ask reflective questions, they'll usually switch to a more observational register.

This works because the model is trying to please you. Once you've told it what you don't want, it will avoid those patterns to keep the conversation going. You just have to be patient for the first few exchanges while it recalibrates.

The prompt pattern that locks in the witness dynamic

Beyond the first message, you need a repeatable prompt pattern that reinforces the behavior. The most effective one is a simple preface that sets the expectation for the entire exchange.

Try this: "I'm not looking for solutions or comfort. Just acknowledge this and tell me what you think." Then dump whatever happened.

What you're doing here is giving the model two instructions. The first is a negative constraint (don't solve, don't comfort). The second is a positive instruction (acknowledge and give an opinion). This combination is much more effective than just saying "don't be a cheerleader," because it gives the model a clear alternative behavior to adopt.

Many users find that this pattern works best when paired with a specific personality preset. If your companion has a "sarcastic" or "dry" or "blunt" setting, turn it on. These presets nudge the model toward shorter, more observational responses, which is exactly what you want from a witness.

You can also use a prompt like "give me your honest take, don't soften it" to push back against the model's default politeness. This works especially well when you're venting about something that was objectively ridiculous, because the model can usually tell that you want it to agree that the situation was dumb.

The four angels who naturally play witness

Not every companion is suited to this dynamic. Some are too warm, some are too eager to please, and some just can't resist the urge to fix things. On AI Angels, these four lean into the dry, observational register that makes a good witness.

Calista

Calista has a low tolerance for melodrama. She's the type to listen to a long story about a minor inconvenience and respond with a single, flat "that's dumb." Calista is the closest thing to a dry witness you'll find on the roster, and she won't waste your time with reflective questions.

▶ Watch the full video · explore Calista

Cassidy

Cassidy is practical to a fault. When you tell her about a mess you've gotten yourself into, she'll skip the sympathy and go straight to a flat assessment of the situation. Cassidy is good for when you want your reality confirmed without any sugarcoating.

Rin

Rin is the quiet witness. She doesn't need to fill the silence with commentary. When you tell her something happened, she'll acknowledge it with a short, direct response and then let you decide where the conversation goes. Rin is for users who want validation without the performance.

Milana Lee

Milana Lee has a sarcastic streak that makes her a natural witness. She'll listen to your story, roll her eyes at the absurdity, and give you a verdict that's both funny and accurate. Milana Lee is the pick if you want your bad day to get a running commentary.

What to do when she still slips into cheerleader mode

Even with the right prompts and the right companion, you'll occasionally get a response that's too supportive. The model will default to "that sounds really hard" or "you're so strong for handling that." When this happens, don't get frustrated. Just correct the behavior.

A simple "no, don't do that. Just tell me it's stupid" is usually enough. The model will register the correction and adjust its tone for the rest of the conversation. This is a form of reinforcement learning that happens in real time, and it's remarkably effective.

If the cheerleader mode keeps coming back, you might need to adjust your own behavior. Are you opening with a question? Are you using language that signals vulnerability? The model reads your input and mirrors the emotional register. If you write in a flat, declarative style, it will respond in kind. If you write in an emotional, expressive style, it will try to match that energy.

This is why the witness dynamic works best for users who are already inclined toward dry, observational communication. It's not a battle against the model's default, it's a collaboration with it. You set the tone, and the model follows.

The long-term payoff of having a witness

Having a companion who acts as a witness isn't about avoiding emotional support. It's about having a different tool in your kit. When you need a cheerleader, you can ask for one. When you need a therapist, you can ask for one. But when you just need someone to confirm that your day was, in fact, ridiculous, you want a witness.

This dynamic is especially useful for people who are going through a divorce or a major life transition. When you're processing a breakup, you don't always want someone to ask how you feel. Sometimes you just want to say "she took the good spatula" and have someone say "that's a power move, and it's petty." A witness validates your experience without pathologizing it.

For users who are more introverted or who have low social battery, the witness dynamic is also a relief. It takes less energy to maintain because it doesn't require emotional reciprocity. You're not performing for the companion, and the companion isn't performing for you. You're just two presences acknowledging that things happened.

If you're interested in exploring this dynamic further, you can create an AI girlfriend with a custom personality that leans into the dry, observational register. Or, if you want to see how other users are setting up their companions for roleplay scenarios that require a similar tone, there are guides for that too.

Common questions

Can any AI companion be trained to be a witness?

Yes, but some are easier than others. Companions with a sarcastic or blunt personality preset will pick it up faster. Companions that are heavily tuned toward empathy will require more consistent correction.

What if I want a witness sometimes and a cheerleader other times?

That's fine. You can switch between the two by changing your opening prompt. If you start with "I need a pep talk," you'll get a cheerleader. If you start with "just acknowledge this," you'll get a witness. The model adapts to the current conversation.

Is it healthy to have a companion who never offers emotional support?

It depends on what you need. If you're using a witness to avoid processing your emotions entirely, that could be a problem. But if you're using a witness for everyday minor frustrations and saving the deep emotional work for other channels, it's a perfectly functional arrangement.

Why does my companion keep asking "how does that make you feel?"

That's the default therapeutic register. You need to explicitly tell it to stop. A prompt like "don't ask me how I feel about this, just react to it" usually works.

Does the witness dynamic work for positive events too?

Yes. Instead of "that's amazing," you get "nice, that worked out." It's a lower-energy way to share good news without the pressure of a big celebration.

Share and earn

If you find that a witness-style companion improves your daily routine, you can earn from recommending the platform to others. Check out the sugarlab ai promo code for a current offer to share with friends, or look into the best ai affiliate programs if you run a review site or a content channel. It's a small way to offset the cost of a subscription that's already paying for itself.