Skip to content

COROS AI · Case Study

Designing an AI that remembers you

My World remembers your COROS AI coaching conversations and shows your breakthroughs, reminders, and open concerns. I designed it end to end and shipped the front end in TypeScript and React on Next.js.

25

customer interviews

20

transcripts per eval batch

5

must-fix issues caught in one run

Prompt EngineeringUX/UIDesign Engineering
Role
Product Designer · Prompt engineering · Front-end
Team
Founder/CEO · 2 engineers · 2 designers
Timeline
April – August 2026, shipped incrementally
Tools
Figma · Claude Code · Streamlit · TypeScript, React on Next.js

The product

COROS AI is an AI coach that helps professionals shift moods, repair relationships and act when they’re stuck; I joined as founding designer in 2025 and led product, UX, prompts, research and brand.

Try what I built
The Coaching Provocation card: 'Identity is built, not discovered,' ending in a question to the user.
A session detail page with summary, entry and exit moods, reminders, and a breakthrough quote.
The My World page: a featured band with the breakthroughs donut, a quote, and provocation cards above the Reminders widget.

The problem

Users could not find their breakthroughs

In 25 customer interviews, one thread kept returning: the coaching lands, but the product never shows it.

I had a difficult relationship with one of my nephews. From COROS, I had breakthroughs there. I can’t find them anymore.

So I gave the conversation structure: sessions, topics, and My World.

Architecture

Sessions, topics, and one home

My World, a topic, then one session.

A session is one conversation. A topic is a recurring situation, like “Conversation with dad.” That rule governs the detection pipeline.

The interface

A page you can scan

Filters live inside widgets.
Mode A browses everything. Mode B filters by dimension.

The breakthrough donut was about to be cut. I merged it with the Latest Breakthrough card and specced the state machine.

Done earns confetti. Delete exists because AI can be wrong.

Decluttering

From a data table to a bento box

An early topics page: a dense, spreadsheet-like list, closer to a CRM than a coaching tool.
Before: a dense table.
After: inline metadata, sticky CTAs, dimension filtering.

Early designs showed topics like a CRM. I pitched a widget-style bento box for the whole page.

The AI

The prompts are the product

The Coaching Provocation widget: a synthesis across recent sessions that ends in a question and a Revisit action.
A provocation synthesized across sessions.

  1. 20-session batch in

  2. Score every field (semantic + LLM judge)

  3. Read a stratified sample

  4. Fix the prompt

Re-run the same 20 sessions

Every card comes from prompts I wrote: session summary (v6), topic summary (v7), topic detection, coaching invitation. An ownership gate limits each prompt to what the user said, because a wrong breakthrough manufactures a false memory.

I test them in 20-transcript batches. One run surfaced 5 must-fix issues, each fixed with one added clause.

Conversations are confidential, so examples are redacted or synthetic. The prompts are company IP.

Takeaway

What I owned end to end

I wrote the doc engineers build from, the prompts, the QA tooling, and increasingly the code.

At a pre-seed startup, the most valuable designer is the one who removes handoffs.