1 / 18
Case study 02 · Jobpilot designer · builder · founder

Cutting the product down
to what's worth building.

A solo-built AI job-search coach — shipped, live, and better for losing two of its five features.

5 features → 3 Shipped & live jobpilot.katexu.com
Kate Xu · Product Designer & Builderkatexu.com

Origin

40+ applications in flight —
my own job search.

Six disconnected tools: a resume editor here, salary data there, a tracking spreadsheet, prep notes scattered across apps. The product idea was the frustration itself:

"Why isn't there one place that coaches you through the whole journey?"

resume editorsalary sitestracking sheet prep notesjob boardscalendar

V1

Five pillars. I was proud of the scope.

01

AI job search

02

Resume optimization

03

Interview story bank

04

Market insights

05

Application tracking

Then reality hit from two directions.

Reality · blow zero

The feature I was proudest of.

AI job search — personalized matching that ran your full profile against live listings, with a reason attached to every match.

The most ambitious build in v1. The technical centerpiece. The feature I showed people first.

The pitch

Your profile × live listings
= a reason for every match

Reality

The reality check.

Blow 1 · user testing

Too slow — and not better.

Processing a profile against listings takes time, and the results weren't better than LinkedIn — whose data moat I can't match.

Blow 2 · unit economics

Burning tokens on every search.

Personalized search is expensive to run. A solo builder on freemium can't carry a feature that costs money on every use.

Expensive. Slow. Worse than the incumbent. Any one is survivable — all three is a verdict.

Method

AI builds whatever you ask.
It never pushes back on scope.

Saying no stayed my job. So instead of asking AI to build, I asked it to evaluate — turning the machine that never questions scope into the scope's harshest critic.

The prompt that reshaped the product

"Rank every feature I have by market uniqueness vs. cost to run. Predict which ones people can already get elsewhere."

Decision

Killing a feature that worked.

My second-proudest feature: the resume editor. AI-assisted, line by line. It shipped, it functioned, users liked it. Nothing about it argued for its own death.

Then the market spoke: Teal — $20M in funding, 4M users — a crowded, well-funded category doing exactly this. A solo builder is always behind.

Working isn't a reason to keep something —
being worth building is.

The hardest features to kill are the ones that function.

Research

Every competitor is strong in 1–2 pillars — weak in the rest.

AI SEARCH
RESUME
STORY BANK
MARKET
TRACKING
LinkedIn
strong
weak
weak
weak
Teal
weak
strong
weak
ok
Salary sites
strong
Whitespace
open
open*
open*
Three whitespaces fell out. * strong alone, unbeatable connected — no one ties them into one loop.

Decision

The cut.

01 · LinkedIn owns the data moat

AI job search

02 · Teal: $20M funding, 4M users

Resume optimization

03 · no one offers structured STAR stories

Interview story bank

04 · woven into coaching, not a destination

Market insights

05 · the spine connecting the other two

Application tracking

Killed the two the market already owned. Kept the three no one else connects.

Sub-story · v2, the atmosphere

The cut decided what to build.
Then: how should it feel?

Walk into a room — you feel something before you read anything. Every interface is a space; its light, materials and noise set a mood before any feature loads. For a product people open on their worst days, the mood is the emotional value. So the v2 redesign started from a moodboard, not a wireframe.

Space

The interface as a room

Atmosphere

Light, material, noise

Mood

The emotional value of the product

Sub-story · where v2 started

The moodboard came before the wireframes.

moodboard · 01drop image here
moodboard · 02drop image here
moodboard · 03drop image here
moodboard · 04drop image here
moodboard · 05drop image here
Fed to Claude Design connected to the live codebase — the machine learns the vibe before it generates a screen.

Sub-story · method

Teaching the machine to feel the room.

An AI can't be told "make it cozy." It can be shown a space and taught to extract what makes the atmosphere — held as a system, while the taste calls stay human.

Moodboard Stylescape Key decisions · human Tokens Build
#FAF6EF · paper · 70
#1A1A1A · ink · 20

Dopamine Bauhaus · 70/20/10

Calm first, then a spark: lower cortisol before asking for effort; dopamine only at real progress.

Sub-story · mood → function

Every response corrects your position.

The mood earns its keep in function. Tracking isn't a filing cabinet — every answer that comes back, rejection or next step, is a data point about your standing. The market feature accumulates them into a continuously corrected position: where you actually are, updated as the market keeps talking.

APPLIED3

Staff PD · fintech
Design Eng · AI
Sr PD · dev tools

INTERVIEW1

Sr PD · platform
Design Eng · AI

CLOSED2

Sr PD · social

Your market position

68 ▲ updated after 1 response
signal: 14 applicationsconfidence rising
Numbers are the dopamine: oversized, tabular, one colour — and the overshoot is reserved for real wins.

Design decision

JD storage.

The interview call comes weeks later — and by then the job posting is gone. Jobpilot saves the JD at application time and builds the prep page around it.

"Save this now — future you will thank you."

Principle

Design for the high-anxiety moment, not the entry moment.

Design decision

No sign-up wall.

The full product works anonymously — job seekers are drowning in logins, and value comes before the ask. Sign-up appears only when it buys the user something: opening your story bank on another device. Build stories on your phone, rehearse from your laptop.

The cost: dual-mode storage — local-first, with a seamless merge on account creation. The user never feels it.

Worth noticing

Here, the design decision and the technical decision were the same decision.

Opportunity

Why these two are the killer features.

The habit loop

JD storage & story bank

Both demand frequent return visits — check, update, rehearse. Not one-shot tools: a loop that builds toward a moment.

The moment AI can't take

The interview itself.

Real-time human interaction. AI can prepare you for it; it can't do it for you.

The opportunity: own the preparation loop that sits closest to the irreplaceable human moment.

Reflection

The product got better
when it got smaller.

AI made it easy to say yes to everything — the hard part was saying no. And the atmosphere chapter is the same lesson from the other side: once the product was small enough to be honest, it could afford to feel like somewhere.

The question I now ask first

Where in this product is the irreplaceable human moment?

How I build

Live, instrumented, from day one.

Figma · source of truth Cursor via MCP Claude API Next.js · deployed

User-testing data loops back in, and AI reviews scope against market uniqueness and running cost before anything gets built. Design keeps the veto.

Kate Xu · designed and built solo · jobpilot.katexu.comkatexu.com