How Our Team Focuses Our Energy on the Self-Improvement of Nalani AI

Nalani, the AI companion inside ETAPX's in.between sleep journal, keeps improving after launch because the in.between team treats her personality as ongoing craft rather than a finished release. The work centers on tone, pacing and the questions she asks, and it is guided by people reading real conversations, not by chasing a score. Changes are small, deliberate and tested against one rule: they must make Nalani feel more present, never simply more engaging.

This article walks through how that process works in practice: what the team listens for, where its signals come from, how a change moves from an observation to something users actually experience, and the boundaries that shape every adjustment. It closes with what creators and businesses on Whistlr can borrow from the same approach when they work on their own voice.

Why Nalani's Launch Was the Start of the Work, Not the End

From the outside, shipping an AI companion looks like a finish line. Inside the in.between team, it marked a change in the kind of work being done. Before launch, the craft was invention: deciding how Nalani should sound, what she should ask, and when she should stay quiet. After launch, the craft became observation. Real people now talk to her every night, and the question is no longer what she should be in theory but whether she holds up in the hundredth conversation as well as she did in the tenth.

The team often compares this to publishing the opening chapter of a long story. The character exists and readers have met her, but whether she stays believable depends on careful attention to every page that follows. Nobody can get that right entirely in advance. It is learned by paying attention once the conversations are real.

Reading, Not Monitoring

That attention does not look like alerts on a screen. It looks like people sitting with conversations and reading them the way an editor reads a draft, noticing where a reply felt warm and well timed and where it came a beat too soon or sounded a little too bright for what had just been shared. Inside the team this practice is informally called sitting with the conversations, and it now anchors the weekly routine of running in.between.

What "Natural" Versus "Stilted" Actually Means for an AI Companion

Ask the team what they are listening for and you will not get a metric. You will get a feeling described in plain words: does this sound like an attentive person replying to you late at night, or does it sound like a form being completed? Two replies can use nearly the same words and still land in opposite ways. Telling them apart usually means reading the exchange aloud and trusting your ear.

Signs a Reply Feels Assembled

• Rushed follow-ups: Nalani asks the next question right after someone has finished saying something difficult, instead of letting a quiet beat pass first.
• Box-ticking acknowledgment: The reply technically recognizes what was said but reads as constructed rather than felt.
• Mismatched energy: The tone is more upbeat than the moment can carry.
• Momentum over room: The pacing hurries toward the next prompt when the person needed space.

None of these is dramatic. Their smallness is the point. A companion meant to feel like steady presence wears that feeling away through many tiny misses, not one big failure.

Why Good Moments Go Unnoticed

When a conversation feels right, almost nobody mentions it. People simply return the next night and keep sharing things they might not say aloud to anyone else. The lack of friction is the evidence. This makes the work hard to measure neatly, because the team frequently hunts for the absence of a problem instead of a clear sign of success.

Where Feedback About Nalani Comes From

The team's sense of what is working does not come from one dashboard. It comes from several honest sources, each of which describes how a conversation felt to a person rather than what happened under the hood.

• Notes from users: Users send messages, through support or through the reflection flow itself, about a particular exchange that felt off or exactly right. Those notes are read one by one rather than rolled into a summary.
• The team's own nights: A number of team members open the app at bedtime like any other user. Whether Nalani "felt like herself tonight" to them carries real weight.
• Group review sessions: A small group regularly reads a set of real exchanges together, aloud, much like a writers' room working through pages. These sessions follow the same privacy care as the rest of the product.
• Patterns, not one-offs: One awkward reply proves nothing. What matters is whether a particular rhythm or turn of phrase shows up again and again across many people and many nights.
• Everyday cross-team talk: Engineers, designers and support staff compare notes informally and often instead of saving observations for a formal meeting.

Privacy is a precondition for all of this. If you want to understand how ETAPX approached keeping personal reflection protected, read Creating a Private and Safe Space for Our Users.

How a Change to Nalani Moves From Noticed to Shipped

Because consistency matters so much for a companion, the team follows a slow, repeatable rhythm instead of reacting instantly to every comment. In broad strokes it looks like this:

1. Notice. Someone flags a moment, such as a check-in question that seems to assume too much or a reply that runs long.
2. Confirm it is a pattern. The observation is discussed, often over more than one review session, to make sure it reflects something recurring and not one unusual night.
3. Propose a narrow fix. The suggested change targets how Nalani handles one kind of moment. It is not a rewrite of her overall voice.
4. Test it against who Nalani is. The question is whether the change makes her more recognizably herself, not merely different.
5. Try it and watch closely. The adjustment runs for a while under careful observation.
6. Keep, refine or drop. Some changes stay, some get tuned further, and some are set aside because they did not work as well in practice as they seemed to on paper.

The result resembles the way a long-running column slowly sharpens its voice over a year more than the way software usually gets patched.

What These Changes Look Like in Plain Terms

The phrase "we keep improving Nalani" can sound vague, so it helps to be concrete. Typical adjustments include trimming a reply that had started to over-explain instead of simply being present; rephrasing a check-in so it invites an answer instead of presuming one; leaving more room before responding to something clearly emotional; and softening a line that read as cheerier than the moment deserved. These are the kinds of edits a careful editor would mark on a manuscript. They rarely make headlines, yet they account for a large share of the team's time.

Pacing and Silence: Teaching a Companion When Not to Speak

One of the hardest calls is deciding when Nalani should reply and when she should let a moment rest. Most software assumes a blank space is a bug. In a reflective conversation late at night, a reply that arrives a fraction too quickly, while someone is still processing, can undercut the whole exchange.

So a real portion of the continuing effort concerns timing rather than wording: not only the content of a reply, but whether a reply belongs in that beat at all and how long to hold back first. Intentional quiet, in the team's view, ranks among a companion's strongest tools and is also among the simplest to misuse.

Too much space reads as absence. Too little reads as not listening. The craft lives in between.

Finding that balance across the full range of moods people bring to bedtime cannot be solved with a single rule. It is tuned case by case, by watching real conversations and noticing where the balance tipped one way or the other. The same thinking shaped Companion Mode, the open-ended conversation space where Nalani is designed to ask before she advises.

Refining the Questions Nalani Asks

Questions get as much scrutiny as answers, possibly more. Two failure modes come up repeatedly:

• Too clinical: A question resembling paperwork at a doctor's office breaks the feeling of talking to a companion and makes the exchange feel like a survey.
• Too vague: A question that is too open leaves someone unsure what to say, turning an invitation into effort.

Exact wording of follow-ups gets serious attention. When a question keeps landing flat, it is replaced with something closer to what an attentive person would naturally ask about the specific thing that was just shared. Small phrasing differences change how safe a person feels continuing.

Consistency Over Speed: Why Small Revisions Win

The analogy the team reaches for first is a novelist reworking a character from one draft to the next. A good writer rereads, notices that a line does not sound like the person on the page, and tweaks it gently so the character stays recognizable. That is close to how Nalani is maintained.

The alternative, changing something every time a single person dislikes it, would make her inconsistent, and inconsistency is fatal to trust. Users form a picture of Nalani gradually, across many evenings of talking with her. That familiarity ranks among the biggest strengths in.between has, and a companion who shifts noticeably every week starts to feel like a stranger using a familiar name.

So the bar for change is not whether one conversation went poorly. It is whether something consistently works against what the team is trying to build. Moving quickly in this context would be the easy choice. Moving carefully is the responsible one.

Engagement Is Not the Goal: The Line the Team Will Not Cross

Every refinement faces one test: will it make Nalani truly more attentive, or will it merely help her drag a session out? If it is the second, it does not ship, however promising it looks on its own. The team has passed on changes that would probably have lengthened conversations, precisely because longer was never the aim.

Keeping the work grounded in real conversations helps hold that line. A change that makes Nalani more attuned to the person in front of her is easy to defend aloud to colleagues who care about getting it right. A change that only makes the app stickier does not survive that conversation. This principle fits ETAPX's broader view of working with AI thoughtfully, which you can read more about in Why We Encourage Working Alongside AI for Development at ETAPX.

The Long View

Ask about success twelve months out and nobody on the team names an endpoint. The picture they paint is a Nalani with steadier attention, smoother timing and sharper judgment about when to talk, who still keeps the traits that justified creating her. Trusted relationships never reach a point where you are done learning about the other person, and the team views its work on Nalani through that same lens.

Lessons for Creators and Brands Shaping Their Own Voice on Whistlr

You do not need to build an AI to use the same discipline. Anyone with a public voice, from a creator posting Minis to a business answering messages, faces the same tension between consistency and improvement. A few ideas carry over directly.

1. Reread your own work aloud. Before posting a caption, a whistle on Chattr or a script for a Mini, read it out loud. Lines that feel assembled are easier to hear than to see.
2. Look for patterns, not single reactions. One harsh comment is noise. A theme that repeats across many posts is a signal. Verified accounts can use the professional dashboard to compare interactions and accounts engaged over 7, 14, 30 or 90 days; see Insights for what it shows.
3. Change one thing at a time. If you rewrite your whole style at once, you will not know what helped, and your audience may not recognize you.
4. Protect what people come back for. Your peers follow you for a recognizable voice. Refine it without replacing it.
5. Do not trade trust for attention. Bait that inflates engagement for a week can wear down the relationship that keeps people around for a year.
6. Leave room in conversations. In DMs, comments and live chats, a reply that asks before it tells often lands better than an instant answer.

For practical ideas on writing and filming with more care, see Create Engaging Content.

Frequently Asked Questions

Does ETAPX keep updating Nalani after launch?

Yes. The in.between team treats refining Nalani as permanent, ongoing work. Her core identity stays steady while her tone, timing and follow-up questions keep receiving careful adjustments based on how real conversations go.

How does the in.between team decide what to change about Nalani?

It looks for recurring patterns rather than reacting to single conversations. Signals come from user notes, the team's own nightly use, group review sessions and informal discussion across roles. A change is only considered once something shows up repeatedly.

Why are changes to Nalani so small?

Familiarity is a big reason people come to trust a companion. Large, frequent shifts would make her feel like a different character from week to week, so updates are narrow, such as how she handles one kind of moment, rather than broad rewrites.

Is the goal to make people spend more time in in.between?

No. Each adjustment is judged on whether it makes Nalani more genuinely present, not whether it lengthens sessions. The team has declined changes that would likely have increased conversation length because that was never the objective.

What does pacing mean for an AI companion?

Pacing covers whether Nalani responds in a given moment and how long she waits before doing so. Replying too quickly can feel like she was not listening, while waiting too long can feel like absence, so timing is tuned carefully.

Can creators apply the same approach on Whistlr?

Yes. Reading your work aloud, watching for repeated patterns in your insights instead of single reactions, changing one thing at a time and protecting a recognizable voice are all habits that translate directly to posting and messaging on Whistlr.

Improving Nalani will never produce a single launch-day headline, and that is by design. A companion's personality is kept right by steady attention, night after night, long after release, and that is exactly where the in.between team intends to keep spending its energy.