All work

Project 012026

NaraAn AI learning companion that plans, explains, and adapts

A study companion that turns AI output into something learners can actually trust and follow.

Role
Product Designer — 0→1
Timeline
4 months
Team
2 engineers, 1 PM, me
Scope
ResearchUXUIPrototypingAI patterns

Context

An early-stage team wanted to build a learning product around LLMs, but the market was already flooded with “AI tutors” that generate content nobody finishes. They asked me to shape the product from zero: positioning, experience, and interface.

My first job wasn't screens. It was figuring out what an AI learning product should actually promise — and to whom.

The problem

Learners don't lack content; they lack a plan they believe in. Generic AI tutors respond to whatever is asked, so the learner carries the cognitive load of deciding what to study, in what order, and whether the answer is even right.

Research interviews with 12 students and self-learners kept circling the same three pains: plans collapse after day three, AI answers feel unaccountable, and progress is invisible.

Problem evidenceWhere the experience breaks

Framing

We reframed the product from “an AI that answers” to “a companion that commits.” Nara would not just generate study material — it would commit to a plan, show its reasoning, and adjust openly when the learner falls behind.

That framing turned trust from a vague principle into three concrete design constraints: always show why the AI suggested something, make the plan editable by the learner, and never hide a change the AI made.

Key decisions

Decision 01

Show the reasoning, not just the answer

Every AI suggestion ships with a one-line rationale (“adding recall practice — you missed 3 of 5 flashcards this week”). Learners override it in one tap, which paradoxically increased trust in suggestions they kept.

Decision 02

A weekly plan as the home, not a chat box

Chat invites unbounded prompting; a plan gives the product an opinion. The home screen is a seven-day plan the learner can drag, swap, or shrink — the AI proposes, the learner disposes.

Decision 03

Calm density over clever empty states

Instead of decorative empty states, every screen answers “what should I do next, and why.” Progress is shown as momentum (streak of completed intentions), not as a guilt-inducing percentage.

The solution

Solution view 01Nara

A plan you can argue with

The weekly plan is the product's spine. Each card carries the AI's rationale, an effort estimate, and a one-tap override — so the learner stays in control without doing the planning work alone.

Solution view 02Nara

Explain-as-you-learn

Inside a session, explanations arrive in layers: a plain answer first, the reasoning behind it on demand, and related gaps the AI noticed — each layer opt-in rather than dumped at once.

Outcome

+38%

week-4 retention in closed beta

4.8/5

“I trust the plan” score, exit survey

72%

of AI suggestions kept unedited

Reflection

The biggest lesson: with AI products, the interface is mostly trust made visible. Every rationale line and override control did more for adoption than any visual polish pass.

It also sharpened how I work with engineers — agreeing on “what the AI is allowed to decide” as a design constraint removed half of our implementation debates before they started.