Hash Health: trust before growth

For a 0→1 AI product, growth is the wrong first question. Instead ask these: Did users reach their first value moment? And do they trust the AI enough to come back? Answering these tells us whether an AI product is earning its place. Onboarding is the gate to both.

That’s the lens I brought to Hash Health. Hash (hashhealth.io) is an AI-powered nutrition tracking, medication safety, and insights app built to help users achieve better health outcomes. The activation moment that mattered was a new user analyzing their first meal.

But the trade-off: every question we asked to personalize the experience was a question standing between the sign-up and that first meal analysis. So, how do we find the right balance?

We launched with chat-based onboarding. Why not? The hypothesis: this is an AI product, and conversational onboarding signals that. Users would share their goals, an assistant would ask follow-ups.

Then two patterns emerged:

1. Talking to the first few users revealed chat-based onboarding felt like doing work before they had reason to do the work. They hadn’t seen what the product could do yet. They were doing setup, not exploring.
2. Engineering cost without clear value. Handling off-script input gracefully was a non-trivial investment for a feature users hadn’t yet found valuable.

I noticed the conversational moment didn’t need to live in onboarding. Hash already had a built-in assistant inside the app.

We stepped back from chat and tried a standard carousel. Why not ask more? The instinct: more context up front meant sharper personalization - name, age, height, weight, sex, medications, health goal, activity level, family member details.

The funnel numbers jumped:

• ~80% of users who started onboarding completed it
• 70% went on to log their first meal analysis

But talking to users revealed a different qualitative signal. The carousel felt long, and while users could skip some steps, even getting past those skips felt like friction. Every extra tap was quietly pushing them further from the moment we were actually trying to unlock: that first meal analysis.

I dug into the data. (Analysis via PostHog MCP, queries in Claude Code.) Less than 5% of users had added a family member. Of the meals analyzed so far, almost none were for anyone other than self. Two signals pointing the same way: we were building for a use case that wasn’t there yet.

The current version is built around one idea: earn the ask.

Stage 1, at signup - just the basics: name, age, sex, medications, health goal. Enough to know who they are and deliver a first layer of personalization before they’ve analyzed a single meal.

Stage 2, an opt-in CTA on the home screen: “Get my daily target.” Available as optional, once they’ve felt the product’s value. It doesn’t gate the core experience: analyzing the first meal.

Everything else (family member details, medications for family) moved silently to profile. As we analyze user behavior, we’ll understand whether and how to bring any of it back.

When designing onboarding, the order of asks matters as much as what is asked. Get the sequence right, and activation follows.

V3 is currently in the works. We have a hypothesis to test, not a result to claim.

Getting a user to their first meal analysis is one layer of trust. Whether they trust the AI’s output (whether the macros are accurate, whether the food-medication interaction warning is right) - that’s a different layer entirely. That’s where AI evals come in: rubrics, confidence scoring, source citations, expert review.

Activation gets them in. Evals earn the right to keep them.

That’s the work we’re focused on at Hash. When that loop closes, growth becomes a real conversation.