Back
SwipePhoto Observer Adblock Green VPN SoftShield VPN Fluid Chat AI Hypurr Wallet Yooca
AI Assistant · Case Study

Fluid Chat AI

I designed a multi-tool AI assistant from scratch — chat, image generation, summaries and prompt packs — then used the data to beat the one thing every AI app gets wrong: the blank box.

Lead Product DesigneriOS · 20243-month build
AI tools
Home
Chat
My role
Lead Product Designer — research, UX, UI & design system
Team
1 PM · 3 engineers
Timeline
3 months · shipped 2024
Platform
iOS · AI assistant · freemium
0%
Created something, session 1
from 51%
8.4%
Trial → paid
weekly & yearly plans
+31%
Day-7 retention
vs blank-chat baseline
4.5
App Store rating
at launch

The challenge

Everyone shipped the same blank chat box

The market was flooded with ChatGPT clones: a cursor, an empty input, and a silent “…now what?” Powerful underneath, paralysing on top. Most people don’t arrive knowing the perfect prompt — they arrive with a vague need and bounce when faced with a blinking line.

My brief: turn raw AI power into something a first-timer can do something with in ten seconds — without ever staring at an empty box.

What it does

One app, many superpowers

💬

Smart Chat

Ask anything, in plain language.

🎨

Text-to-Image

Generate art from a description.

📄

Summaries

Web pages & PDFs, distilled.

🧠

AI Assistants

Specialised helpers for real tasks.

Prompt Packs

Rewrite, translate, and more.

⌨️

Code Helper

Write & explain code.

Discovery

Why people bounce off AI apps

I don’t know what to type

The blank box is a wall. Without a starting point, people open the app, stall, and close it.

What can this even do?

Capabilities were invisible. Users had no idea image gen, summaries or assistants existed.

Why would I pay?

Without feeling the value first, the paywall landed as a tax — not an upgrade.

From 10 first-use sessions, a teardown of 8 AI assistant apps, and funnel analysis of where new users dropped before their first result.

Key decisions

Designed from scratch, then tuned on the data

I shipped it, watched where first-timers stalled, and redesigned those moments. Every call is mine — and each one moved a number.

Tool-first home
AI tools
01 Beating the blank box

A tool-first home, not a cursor

What I designed
Instead of opening to an empty chat, I opened to choices — AI Tools, Assistants and Prompt packs as vivid gradient cards. Every tile is a ten-second starting point.
What the data showed
My first blank-chat build leaked hard — only 51% of new users produced anything in session one.
The result
Leading with tools pushed first-session creation to 83% — people did something instead of staring.
Prompt packs
02 Removing the prompt skill

Assistants & prompt packs do the wording

What I designed
Specialised assistants (Tech Helper, Entertainment Advisor) and one-tap prompt packs (Rewrite, Translate) so people never have to know “prompt engineering” to get a great result.
What the data showed
Good results required good prompts — a skill most users simply didn’t have, so outputs felt underwhelming.
The result
Pre-built intents raised output quality and kept people coming back — the biggest lift in day-7 retention.
Chat with image gen
Paywall
03 Value before the ask

Let them feel it, then offer Pro

What I designed
A credit system that lets new users create something real — even generate an image — before the paywall, which then frames Pro as “more of what you just loved,” not a gate.
What the data showed
My early upfront paywall converted poorly — people hadn’t felt the value, so paying felt like a tax.
The result
Value-first framing lifted trial→paid to 8.4% — strong for a consumer AI app.

The hardest problem in AI design isn’t the model. It’s the empty input field.

Impact

What shipped — and what it moved

83%
Created in session 1
51% → 83%
8.4%
Trial → paid
value-first paywall
+31%
Day-7 retention
vs blank-chat baseline
4.5
App Store rating
at launch

Measured over the first 6 weeks via Amplitude and App Store Connect. The tool-first home and prompt packs drove the biggest gains — getting people to a result before the blank box could stop them.

Reflection

What I’d do differently

🎛️

Tool-first can hide the open chat

A few power users wanted to just type. I’d keep the tools but make a plain conversation one tap away from anywhere.

🪙

Credits confused first-timers

“+11” wasn’t self-explanatory. I’d show what a credit buys at the moment it’s spent, not just a number.

🌈

The fluid gradients fought the content

Beautiful brand, but busy behind dense chat. I’d reserve the loud gradients for moments, not every surface.

Design system

Fluid, vivid, alive

Palette & gradient

Magenta
#E91E8C
Pink
#FF6EB4
Purple
#8B3DFF
Teal
#2EC5CE
Fluid
gradient

Typography — Inter display, DM Sans body, Space Mono accents

Display · 900
Fluid AI
Heading · 800
AI Tools
Body · 400
Psychedelic concept art with rainbows and bright colors.
Mono · meta
PRO · +11 credits

Key components

Text-to-Image

Tool Card

Vivid gradient entry points — the home of the app.

Make it vibrant with a touch of pastel colors

Chat Bubble

Clean dark bubbles over the fluid background.

Send · +11

Send / Credits

The gradient action, with credit context.