TL;DR: Onboarding has one job: get a new user to the first moment the app is clearly worth it, as fast as possible. Everything else is a cost. My rules for a small app: define the single "first value" moment before designing any screen; ask only the questions that change what the user sees next; let people use the product before you make them create an account where the platform allows it; put permission prompts and the paywall after that first value, not before it; design the empty state as carefully as the full one; and measure drop-off step by step so you cut screens based on evidence, not taste. A short onboarding is not the goal. A short path to value is.
When LectureNotes AI, the note-taking app I co-founded with Benjamin Rhodes, picked up 15 million organic Instagram views in two months, the lesson I took away was not about the videos. It was that reach and conversion are separate problems. A stranger who taps through from a reel has a few seconds of curiosity and no loyalty yet. Whatever stands between them and their first set of good notes is where you lose them. I have since applied the same thinking on Lifemaxxing AI, the self-improvement app I build with Benjamin Siegel, and on client apps. This post is the approach I use now.
Start by naming the first value moment
Before drawing a single onboarding screen, write one sentence: "A new user knows this app is worth it when ___."
For a lecture note-taking app, it is when they see a clean summary and outline of something they recorded. For a plan-and-tracking app, it is when they see a plan that is clearly about them, not a generic template. For a sports app with live leaderboards, it is when their first session shows up ranked against other people.
Once that sentence exists, every onboarding screen gets one question: does this get the user closer to that moment, or is it in the way? Most screens that teams add to onboarding are in the way. They exist because someone wanted a feature tour, a mission statement or a data point for marketing.
Ask only the questions that change what happens next
Personalisation questions are useful when the answers change what the user sees. They are friction when they do not.
On Lifemaxxing AI, the questions earn their place because the product is a personalised plan, not a chatbot. What the user says about their goals and routine is the input that makes the first plan feel like theirs. Skipping those questions would make the first value moment worse.
The test I apply to every question:
- Does the answer change the first screen of real product the user sees? If yes, keep it.
- Is it only for analytics or marketing? Move it later, or infer it from behaviour. You can learn a lot from what people actually do in the first week without asking.
- Can it have a sensible default? If most people would pick the same answer, set the default and let them change it in settings.
Show progress honestly ("3 of 5"). And when the answers feed an AI step, show the result being built from them, so the user sees the connection between what they said and what they got. That connection is a big part of why a personalised plan feels personal.
Let people try it before the account wall
Account creation is the single most common place I see new users disappear. Every field, every email verification step and every "check your inbox" is a chance for someone curious to decide it is not worth it.
Where the product allows it, I prefer this order:
- Use first. Let the user reach the first value moment on the device without an account, with their data stored locally or under an anonymous session.
- Save second. Ask them to create an account when there is something worth keeping: "Sign in to save these notes" lands very differently from "Sign up to continue" on launch.
- Make sign-in one tap. Sign in with Apple and Google cover most people. Email with a magic link covers the rest. Passwords on mobile are a last resort.
The engineering detail that makes this work is linking the anonymous session to the new account so nothing is lost. Firebase Authentication, which I use on many projects, supports upgrading an anonymous user to a permanent one. Whatever you use, test the path where a user who already has an account signs in from an anonymous session with data on it. That merge case is where bugs hide.
Products with payments, health data or shared workspaces may need an account up front, and that is fine. The point is to make the account a deliberate decision, not a default.
Permissions and the paywall go after value
Microphone, notifications, location and photos each come with a system prompt you mostly get one shot at. Asking for all of them on the first screens, before the user knows why, gets a lot of "Don't Allow" that is hard to undo.
The rule I follow is to ask for each permission at the moment it is needed, with a short in-app explanation first:
- Microphone when the user taps record for the first time, not on launch. On a recording app this is the very first action, so the context is obvious. If you are building that kind of product, recording long audio reliably is its own set of problems.
- Notifications after the first value moment, tied to a concrete benefit: "Want a ping when your next notes are ready?" I went deeper on this in push notifications for a small app.
- Anything optional only when the user reaches the feature that uses it.
The paywall follows the same logic. A paywall shown before the user has felt any value asks them to pay for a promise. A paywall shown right after the first good result asks them to pay for something they just saw work. Exactly where it sits depends on your costs and your free tier. On an AI product every free action has a real inference cost, which I covered in AI app unit economics, so a generous free tier is not always affordable. But even a tight free tier can usually afford one real result before the ask. The plumbing behind the paywall is a separate topic, covered in in-app subscriptions and entitlements.
Design the empty state and the waiting state
Two screens decide a lot of first sessions and rarely get design attention.
The empty state
A new user lands on a home screen with nothing in it. An empty list with "No notes yet" is accurate and useless. A good empty state does one thing: it points to the action that produces the first value. One clear button, a single line saying what will happen, and if it helps, an example of what a finished result looks like.
The waiting state
AI products often have a gap between the user's action and the result. Transcribing and summarising a long lecture is not instant. That gap is the most fragile point in the whole first session, because the user has done their part and now has nothing to do.
What helps:
- Say what is happening and roughly how long it takes. A real progress indicator beats a spinner. A spinner with no end feels broken after twenty seconds.
- Let them leave. Tell them the result will be waiting, and send a notification when it is ready if they allowed it. A user who closes the app and comes back to a finished result has a better first impression than one who stared at a loading screen.
- Show something early if you can. A partial result, a title or the first section arriving before the rest makes the wait feel productive.
- Handle failure clearly. If processing fails on a first attempt, say so plainly and offer a retry. Silent failure on the first try is close to a guaranteed lost user.
Measure the funnel step by step
You cannot fix onboarding by opinion. Every team has a theory about which screen is the problem, and the theory is often wrong.
I instrument onboarding with the same approach I described in product analytics for a small app: one event per meaningful step, with a step property, plus one event for the first value moment itself. Then the funnel answers the questions that matter:
- Where do people drop? The step with the biggest fall is the first thing to look at.
- How long does it take to reach first value? Median time from install to that moment is the number I care about most. If it is going up, onboarding is getting worse even if every individual screen looks fine.
- Does reaching first value predict retention? Compare users who reached it in the first session with those who did not. If the gap is large, which it usually is, every improvement to the first session is worth more than a new feature.
Split all of this by platform and by app version. An onboarding step that quietly broke on one Android version will show up as a cliff in one segment long before anyone reports it. Pair that with error monitoring and alerting so a broken first-run path pages someone rather than waiting for the weekly review.
Change it safely
Every new user goes through onboarding, so a bad change hits all of them at once. I put onboarding variants behind a feature flag or remote config, roll a new flow out to a share of new users, compare it against the old one and change one thing at a time. And I keep cutting: the same instinct that decides what to cut from an MVP applies here.
The checklist I use
Before shipping or reworking an onboarding flow, I check:
- The first value moment is written down in one sentence.
- Every question changes what the user sees next, or it has been moved or removed.
- Users can reach first value before creating an account, where the product allows it, and anonymous data survives sign-up.
- Each permission is asked at the moment it is needed, behind a short in-app explanation.
- The paywall appears after the user has seen at least one real result.
- The empty state points to one clear first action, and the waiting state shows progress and lets the user leave.
- Each step and the first value moment are tracked, split by platform and version.
- Onboarding changes ship behind a flag and can be rolled back without a release.
None of this needs a big team. It needs someone who is willing to delete screens.
I build mobile apps, AI products and the backends behind them for startups in the US, UK and Europe, and I care a lot about the first five minutes a new user spends in them. If you are launching an app, or your installs are not turning into active users, read more about my background or book a call.