How We Improve Nalani AI After Launch: Close Reading, Small Revisions and No Engagement Tricks

We improve Nalani AI the way an editor improves a manuscript: by reading real conversations closely, looking for patterns that repeat across many nights, and making small, deliberate revisions to her tone, her pacing and the questions she asks. We don't chase a score, and we never tune Nalani to keep you talking longer. Since Nalani launched inside in.between, our AI sleep and reflection journal, this ongoing refinement has become one of the most regular parts of how the in.between team works. This is a look at that process from the inside: what we listen for, where the signal comes from, how a change gets made, and what we've decided we'll never optimize for.

Why Nalani's launch was the start of the work, not the end

From the outside, launch day looked like a finish line. From the inside, it felt like the day the real work began.

Building a first version of a companion is one kind of craft. You decide what she sounds like, what she asks and when she goes quiet. Living with that companion once real people are talking to her every night is a different craft, and it doesn't stop. The character exists now. People are meeting her. Whether she holds up, whether she still feels right in the hundredth conversation the way she did in the tenth, is something you only learn by paying attention afterward. You can't get all of it right in advance.

That attention takes an unglamorous shape. It isn't a dashboard flashing alerts. It's people on the team reading how conversations actually go: the ones that felt warm and well-paced, and the ones that felt a beat too slow, a question too pointed, or a response a little too chipper for what someone had just shared. Informally, the team calls this "sitting with the conversations." It's closer to how an editor reads pages than to how a system monitors a server.

People ask when Nalani will be finished. We don't think a companion is ever finished, any more than a person you know well is. What we owe people is ongoing attention, every week.

— The in.between team

What makes an AI companion sound stilted or natural?

Ask anyone on the team what they're listening for and you won't get a number. You'll get a feeling, described in ordinary words: does this read like something a genuinely attentive person would say back to you at 11 p.m., or does it read like a form being filled in? Two replies can use nearly identical words and still land completely differently. Telling them apart is mostly a matter of reading them out loud and trusting your ear.

None of the stilted examples are dramatic failures. They're small, and that's why they matter. A companion that's supposed to feel like presence loses that feeling in small increments, not in big ones.

A natural moment, meanwhile, is close to invisible when it's working. Nobody writes in to say "that felt natural." They just keep opening in.between the next night and keep telling Nalani things they might not say aloud to another person. The absence of friction is the signal, which makes the work hard to measure neatly. A lot of the time, the team is looking at negative space: what didn't feel wrong.

Where the feedback that improves Nalani AI comes from

"Feedback" can mean almost anything, so here are the specific sources the team relies on.

• Direct notes from users. People write in, sometimes through support channels and sometimes through the reflection prompts themselves, about a specific exchange that felt off or exactly right. Those notes are read individually, not summarized away.
• The team's own nightly use. Several people on the in.between team use the product before bed like anyone else, and their sense of whether tonight felt like Nalani carries real weight.
• Conversation review sessions. A small group regularly sits with a set of real exchanges, handled with the same privacy care as everything else in the product, and reads them together out loud, the way a writers' room reads pages.
• Patterns across many nights. One awkward exchange proves nothing. What matters is whether a certain kind of pacing or response keeps showing up across many different people and many different nights.
• Everyday conversation across the team. Engineers, designers and support staff talk to each other about what they're noticing, informally and often, without waiting for a scheduled review.

What ties these together is that they're all about how a conversation felt to a person. The team isn't chasing a metric. It's chasing a feeling, described consistently enough by enough different people that it starts to look like a real pattern rather than one person's particular night.

How a change to Nalani actually gets made

The comparison we reach for most is a novelist revising a character across drafts. A writer doesn't nail a character's voice in the first draft and never touch it again. On a reread, a line feels off for who this person is supposed to be, and they adjust it carefully, without turning the character into someone else. That's very close to what happens with Nalani.

In practice, the rhythm looks like this:

1. Something gets noticed: a question landing a beat early, or a tone reading brighter than the moment called for.
2. It gets discussed, often across more than one review session, to confirm it's a real pattern and not a one-off.
3. A narrow change is proposed. Usually that means how Nalani responds in one particular kind of moment, not a rewrite of her voice.
4. The change is tried and watched closely for a while.
5. It's kept, refined further, or set aside if it didn't hold up the way it seemed it would on paper.

The bar for a change isn't "one person disliked this." It's "this consistently works against what we're trying to build." Changes are weighed against the sense of who Nalani is supposed to be, and only made when the team is confident they make her more herself, not less.

Tuning when Nalani speaks and when she stays quiet

One of the hardest things to get right is whether Nalani should say anything at all in a given beat. Most software is built to always respond. A blank space reads, to a lot of product thinking, as a bug rather than a choice.

In a reflective conversation late at night, a reply that arrives half a second too fast, on top of something a person was still processing, can undercut the whole moment. So a meaningful share of the ongoing work is about pacing rather than content. What Nalani says matters, but so does whether she speaks at all and how much space she leaves first.

The team has come to think of deliberate silence as one of the most powerful tools a companion has, and one of the easiest to get wrong in either direction. Too much space reads as absence. Too little reads as not really listening. Getting that balance right across the huge range of moods people bring to a conversation before bed is something you can only tune by watching real conversations and noticing, case by case, where it tipped.

Rewriting the questions Nalani asks

The questions Nalani asks get as much attention as her responses, possibly more. A question that's too clinical, anything that starts sounding like an intake form, breaks the feeling of talking with a companion rather than completing a survey. A question that's too vague leaves people unsure what to say back, turning a moment that should feel inviting into one that feels like effort.

So the team spends real time on the exact phrasing of follow-up questions. Ones that consistently land flat get revisited and replaced with versions that sound like something an attentive person would actually ask, in that moment, about the thing someone just shared.

Examples of how we improve Nalani AI's responses

"We're always improving Nalani" can sound vague enough to mean nothing, so here's what the adjustments tend to look like in plain terms:

• Shortening a response that had started explaining too much instead of simply being present.
• Rewording a check-in question that a number of people found slightly presumptuous, so it invites rather than assumes.
• Adding more room before Nalani responds to something clearly emotional, so the pacing matches the weight of what was shared.
• Softening a phrase that, read back later, came across as more cheerful than the moment deserved.

None of these would make a press release. They're the kind of thing a careful editor flags in the margin. That's why this work is easy to overlook, even though it takes up a meaningful share of the team's time.

Spend an afternoon reading conversations back to back and you start to feel it when a reply is half a beat too fast. Most of the work turns out to be restraint: cutting a sentence, not adding one.

— The in.between team

Why we change Nalani slowly, on purpose

It would be faster, in a narrow sense, to make big changes whenever something looks off. We've deliberately chosen not to, even though moving fast is a strong instinct in software.

A companion that changes noticeably from week to week stops feeling like someone you know and starts feeling like a stranger wearing a familiar name. People build a real sense of who Nalani is over weeks and months of nightly conversations, and that familiarity is one of the most valuable things in.between offers. Protecting it means resisting the urge to overcorrect every time a single conversation feels off.

So the pace is closer to how a magazine might slowly refine a columnist's voice over a year of issues than to how software usually gets patched. Narrow, considered and watched.

What we will never tune Nalani for

There's a boundary the team keeps in view through all of this. Refining Nalani's tone and pacing is never about making her more persuasive, more "engaging" or better at keeping someone talking longer than they meant to. That would be a different kind of optimization entirely, and it's not the one happening here.

Every adjustment gets weighed against one question: does this make Nalani feel more genuinely present and attentive, or does it just make her better at extending a session? If it's the second, it doesn't ship, however promising it looks on its own. The team has turned down changes that would likely have made conversations longer, specifically because longer was never the point.

That's part of why the work stays rooted in reading real conversations rather than chasing an abstract idea of "better." A change that makes Nalani more attuned to the actual person in front of her is easy to defend out loud, in a room, to people who care about getting it right. A change that only makes the product stickier doesn't pass that test.

What this means for you as an in.between user

If you use in.between every night, most of this should be invisible to you, and that's by design. A few practical things are worth knowing.

• Nalani should feel consistent. Changes are narrow and gradual, so she stays recognizably herself from one night to the next.
• Your notes matter. If an exchange feels off, or exactly right, telling us through support is one of the sources the team reads individually.
• Shorter is fine. Because we don't tune for session length, a brief night with Nalani is exactly as valid as a long one.
• Improvements show up as feel, not features. You're more likely to notice a conversation sitting a little better than to spot a new button.

Where Nalani's refinement goes from here

Ask the team what success looks like further out and nobody describes a finish line. What they describe is a Nalani who keeps feeling more like herself: more consistently attentive, more reliably well-paced, better at knowing when to speak and when to simply be there, without losing the qualities that made her worth building.

It's an odd goal, because it has no clean endpoint. But it fits what a companion is supposed to be. You don't finish getting to know someone you trust. You keep noticing new things about them, and the relationship stays worth tending because of that. That's the frame the in.between team works in now: a character who keeps being written, carefully, by people still paying close attention to how she sounds at the end of someone's long day.

Frequently asked questions

Does Nalani AI keep improving after launch?

Yes. The in.between team treats refining Nalani as ongoing work, not a phase that eventually ends. The core of who Nalani is stays consistent, while her tone, pacing and the way she phrases questions keep getting careful attention based on how real conversations go. The approach is closer to a writer revising a character across drafts than to shipping periodic software patches.

How does the team decide what to change about Nalani?

Mostly by noticing patterns across many conversations rather than reacting to a single one. The team reads real exchanges in review sessions, reads direct notes from users individually, and uses its own nightly use as a gut check. A change is only made when a certain kind of pacing or phrasing keeps working against what Nalani is meant to be, across many different nights.

Is Nalani being tuned to keep me using in.between longer?

No. That is explicitly not the goal, and the team has turned down changes that would probably have lengthened conversations for exactly that reason. Every adjustment is judged by whether it makes Nalani more genuinely present and attentive, not whether it makes the app stickier. Anything that reads as optimizing for engagement doesn't move forward.

Why does Nalani change in small steps rather than big updates?

Because familiarity matters. Part of what makes a companion feel like a companion is being recognizably herself from night to night, and a voice that shifted dramatically and often would undermine that trust. So changes tend to be narrow, such as how Nalani responds in one specific kind of moment, and they're tried and watched before being kept.

Can I give feedback on a conversation with Nalani?

Yes. People send notes about specific exchanges, through support channels and sometimes through the reflection prompts themselves, and the in.between team reads those notes individually rather than summarizing them away. Feedback about a moment that felt off, or one that felt exactly right, is one of the main signals that shapes how Nalani's tone and pacing are refined.