From Interrogation to Creative Session

The Situation

Stampy is HeartStamp's AI agent — the primary way users interact with the product. It tours users around the site, helps them find the right greeting card, co-writes the card's message, and generates custom artwork. If Stampy fails, the product fails.

And the data said Stampy was failing. To gather enough context to produce a customized greeting card, Stampy needed a minimum of 5 questions, as many as 11, and an average of 8. Worse, some of those questions were genuinely hard to answer. "What tone of message do you want?" and "What art style do you have in mind?" are questions most people walking into a card shop couldn't answer either — they know it when they see it.

The result: the experience felt like an interrogation, and 96% of users bounced before ever seeing a card rendered. The magic of the platform — watching a custom card come to life — was locked behind a wall of questions almost nobody got through.

What I Did

The team's instinct was to optimize the questions — reword them, reorder them, trim one or two. I stepped back and rethought the interaction model itself. The insight: we didn't need users to answer our data requirements. We needed to extract the same information from choices users would enjoy making.

I redesigned the flow into four steps:

1. "What's the occasion, and who is it for?" — The one question every visitor already has an answer to; it's why they came. Two data points in one easy step.

2. "What does [recipient] like to do, and what's the vibe of your relationship?" — Again two data points in one step, and again something the user actually knows. Nobody struggles to describe their sister's hobbies.

3. Pick from 4 message concepts. Instead of asking "what tone do you want?", Stampy proposes four card concepts spanning a range of tones — playful, funny, romantic, solemn, respectful — generated from the first two answers. The user just picks the one that feels right. Their selection tells us the tone and the nature of the relationship without ever asking. The interrogation becomes a creative session: instead of being quizzed, the user is choosing between interesting options.

4. Pick an art style from real cards. Rather than asking users to name an aesthetic, Stampy shows examples of real cards in styles appropriate to everything learned so far. "I know it when I see it" becomes the mechanic instead of the obstacle.

Same data requirements. Completely different experience. My lead AI engineer implemented the redesign; I owned the interaction model, the specs, and the acceptance criteria.

The Result

Completion went from roughly 4% (a 96% bounce rate before users ever saw a card) to 87% of users making it through all four steps and experiencing the platform's magic moment — watching their custom card get created.

What This Proves

First: I treat AI agent design as a product problem, not a prompting problem. The fix wasn't a better system prompt — it was recognizing that the interaction model itself was wrong, and that users shouldn't be asked questions they can't comfortably answer.

Second: I know how to convert data requirements into user delight. Every piece of context Stampy needed was still collected — but through choices that felt creative rather than interrogative.

Third: this is what my product leadership looks like in practice — I found the problem in the funnel data, redesigned the experience, wrote the specs and acceptance criteria, and handed a buildable design to engineering. Concept to shipped feature, across every function it touched.