AI Gardening App

An AI-powered learning platform that turns nonfiction topics into interactive, branching learning experiences with sourced information.

AI Product Design ยท Information Architecture ยท Agentic Workflows

Four screens from the card-collection flow: a check-in card on Home, one question, cards to save, and saved cards two days later.
Four screens from the card-collection flow: a check-in card on Home, one question, cards to save, and saved cards two days later.

WHAT I DID

Designed the interface between gardener and agent.

For the team, I turned user research into low-friction logging (widgets, check-ins, voice and photo capture) and implemented AI workflows that sped up our path from concept to annotated flows.

TEAM

1 stakeholder, 1 product manager, 3 designers, 4 data scientists

METHODS

User Stories, Information Architecture, Prototyping

TOOLS

Figma, FigJam, Claude

TEAM

1 stakeholder, 1 product manager, 3 designers, 4 data scientists

METHODS

User Stories, Information Architecture, Prototyping

METHODS

User Stories, Information Architecture, Prototyping

TOOLS

Figma, FigJam, Claude

THE PROBLEM

An empty agent cannot give valuable advice.

Unlogged daily events cause the digital record to fall behind the physical garden, making the agent's advice obsolete. Only the gardener knows exact harvests or small changes, so the hard part is getting that data in.

Illustration: Growzi's garden record stays flat while the real garden drops with each unlogged harvest, opening a gap between them.

THE CHALLENGE

Designing for dirty hands and clean data.

If everyday events go unlogged, the agentโ€™s advice falls behind. The goal was to make logging quick and inviting, so small check-ins keep the record matched to the real garden.

THE RELATIONSHIPS

Designing for dirty hands and clean data.

If everyday events go unlogged, the agentโ€™s advice falls behind. The goal was to make logging quick and inviting, so small check-ins keep the record matched to the real garden.

Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.
Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.

What the gardener sees

Everything in the moment, but none of it written down.

Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.
Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.

Growzi's record

The agent reasons over this record and outside data like weather. It can estimate, but it canโ€™t observe.

THE LOOP

Tightening the loop through progressive inputs.

The system moves through four phases: establishing context via home screen check-ins, resolving ambiguity through processed natural language, predicting outcomes with updated forecasts, and guiding action when insights are resurfaced.

01.

Establish context

Growziโ€™s agent notices nothingโ€™s been logged during a busy stretch and asks for a quick look.

02.

Resolve ambiguity

The gardener answers in one sentence. The agent plays back what it heard (โ€œAbout 8โ€“10 lbs you didnโ€™t logโ€) so nothing is guessed.

03.

Predict outcomes

The agent updates the forecast: 4โ€“5 days of prime fruit left. The gardener saves it for later.

04.

Guide action

Two days later, the saved card is still current, and โ€œPlan my givesโ€ turns it into action.

PRINCIPLE 01

Log in passing.

Growziโ€™s agent takes input in small pieces. The home screen widget lowers the data-entry threshold to a single tap or a raw verbal sentence, capturing context immediately without requiring an open app session.

Three steps: a lock-screen widget asks "Picked anything today?", the gardener sends a one-sentence voice note, and it is saved to the fruit tree's harvest log.
Three steps: a lock-screen widget asks "Picked anything today?", the gardener sends a one-sentence voice note, and it is saved to the fruit tree's harvest log.

PRINCIPLE 02

Ask when it matters.

Gardeners won't remember to log, and the agent can't wait forever. When the agent notices nothing's been logged during a busy stretch, it asks through a check-in card on Home. It fills in what it can, asks one question at a time for what's missing, and nothing is logged until the gardener confirms.

Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.
Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.

Isolate attention

Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.
Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.

Protect focus

Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.
Three annotated screens: a check-in card on Home, one question answered with a tap, and the agent playing back its rough estimate before anything is logged.

Guarantee data integrity

PRINCIPLE 03

Make responding feel low-stakes.

Responding is always optional. Cards can be dismissed and wait until the gardener has a free moment.

TAKEAWAYS

What I learned about designing for AI-native products.

01

Designing for AI means designing a relationship

An agent is only as good as what people tell it, and the interface decides whether sharing feels like a chore or comes naturally.

02

The stakes are higher for data integrity.

Every interaction feeds the model the agent learns from. A confusing screen frustrates users and feeds the agent bad data.

03

Traditional UX fundamentals still apply.

Systems still need to match how people already think.

TAKEAWAYS

What I learned about designing for AI-native products.

01

Designing for AI means designing a relationship

An agent is only as good as what people tell it, and the interface decides whether sharing feels like a chore or comes naturally.

02

The stakes are higher for data integrity.

Every interaction feeds the model the agent learns from. A confusing screen frustrates users and feeds the agent bad data.

03

Traditional UX fundamentals still apply.

Systems still need to match how people already think.

TAKEAWAYS

What I learned about designing for AI-native products.

01

Designing for AI means designing a relationship

An agent is only as good as what people tell it, and the interface decides whether sharing feels like a chore or comes naturally.

02

The stakes are higher for data integrity.

Every interaction feeds the model the agent learns from. A confusing screen frustrates users and feeds the agent bad data.

03

Traditional UX fundamentals still apply.

Systems still need to match how people already think.

REFLECTION

How our early-stage team explored with AI.

An AI-native app means designing how people work alongside an agent. Our early-stage team moved by trial and error, and AI tools let us explore many directions fast.