Redesigning Kaze AI's core creation experience — retention, scale, and new AI capabilities.
ABOUT KAZE
Kaze is an AI creative tools startup focused on China and US markets. It grew its early user base through utility features like watermark removal — high traffic, low retention. I joined as Lead Designer in September 2025, and spent the next eight months helping the team transition into a platform people come back to. This case focuses on Create, which was at the center of that shift.
From one-off generation to continuous creation
Problem
Before the redesign, Chat Edit was a single-purpose image generation tool — one prompt, a batch of results, done. It was the highest-traffic feature on the site. But most users came once, generated something, and left. Retention was low, repurchase rates were low, and the product had no mechanism to change that.
The question was: how do we turn this into something users actually build a habit around?
Narrowing to the target user group
Discovery
Prompt complexity turned out to be a useful signal. Most users kept things simple and left. But a smaller group — more complex prompts, clearer use cases, higher spend, better retention — showed what the product could be for. Professional users, especially in Western markets, had meaningfully higher repurchase rates and brand influence. The business case was clear.
We narrowed focus to US and European e-commerce sellers and content creators and went deeper with that group. The gaps were consistent: no continuity between sessions, a capability ceiling they'd already hit, and enough friction that the underlying ask — one tool, picks up where I left off — wasn't close to being met.
The platform lacked a complete creation loop. For this segment, that wasn't a minor inconvenience — it was the reason they weren't staying.
Three moves toward a continuous creation loop.
Strategy
The redesign wasn't just about fixing the creation screen — it was about building the full context around it.
01
Chat Edit became Create — repositioned as the primary entry point, with room to absorb new capabilities without fragmenting the product.
02
The barrier to start and re-create came down — presets, saved assets, and AI suggestions so the second session feels as easy as the first.
03
Create became the center of a larger loop — homepage inspiration, asset management, personal profile — turning a single good session into a reason to come back.
The design philosophy across all three: every moment after the first generation should feel like a natural next step, not a fresh start.
Designing for three moments:
Getting started, Coming back, Staying in flow.
3 Solutions
The original layout stacked prompt above results. I moved to a horizontal split — creation on the left, results Repositioned from one tool among many to the primary entry point. Users land somewhere with a clear purpose, not a list of utilities to pick from.
01
IA shift: Chat Edit → Create
Repositioned from one tool among many to the primary entry point. Users land somewhere with a clear purpose, not a list of utilities to pick from.
02
Preset library + AI prompt suggestions
For users who know what they want but don't know how to say it, presets and AI suggestions bridge the gap. The empty prompt box is the hardest moment — these reduce that friction before the first generation.
03
Gallery as inspiration, not just showcase
Gallery content lives inside the creation flow — not as a separate destination. Browsing and creating happen in the same context, so inspiration leads directly to action.
Getting started, Coming back, Staying in flow.
The original layout stacked prompt above results. I moved to a horizontal split — creation on the left, results Repositioned from one tool among many to the primary entry point. Users land somewhere with a clear purpose, not a list of utilities to pick from.
01
IA shift: Chat Edit → Create
Repositioned from one tool among many to the primary entry point. Users land somewhere with a clear purpose, not a list of utilities to pick from.
02
Preset library + AI prompt suggestions
For users who know what they want but don't know how to say it, presets and AI suggestions bridge the gap. The empty prompt box is the hardest moment — these reduce that friction before the first generation.
03
Gallery as inspiration, not just showcase
Gallery content lives inside the creation flow — not as a separate destination. Browsing and creating happen in the same context, so inspiration leads directly to action.
Dive: lists that let you think.
List view
The old list view was a filtering treadmill. Users spent most of their time adjusting filters, refreshing, and hunting for signals the system wasn't surfacing. Horizontal scrolling meant you could never see the full picture at once.
The redesign was built around multi-thread monitoring — the actual mental model of dispatchers and coordinators. Orders are now grouped by relationship and relevance. Inline risk tags surface anomalies within the flow rather than requiring a separate check. Quick-action buttons appear contextually based on order status.
We also reduced the default filter set based on usage data — showing only what people actually used, with an option to customize further. This alone cut decision overhead significantly.
Focus: detail pages that
guide instead of overwhelm.
List view
The old list view was a filtering treadmill. Users spent most of their time adjusting filters, refreshing, and hunting for signals the system wasn't surfacing. Horizontal scrolling meant you could never see the full picture at once.
The redesign was built around multi-thread monitoring — the actual mental model of dispatchers and coordinators. Orders are now grouped by relationship and relevance. Inline risk tags surface anomalies within the flow rather than requiring a separate check. Quick-action buttons appear contextually based on order status.
We also reduced the default filter set based on usage data — showing only what people actually used, with an option to customize further. This alone cut decision overhead significantly.
What shipped,
and what it proved.
Impact
The rollout began with a gray-release pilot. Frontline staff tested the new system alongside the old one.
Dashboard
+20%
Faster core workflows
Dashboard
+20%
Faster core workflows
Dashboard
+20%
Faster core workflows
Dashboard
+20%
Faster core workflows
"The new system looks great and makes my job much easier. I no longer need to keep adjusting filters."
Chen Xiaochu — Dispatcher, 13 years experience
""Every screen clearly tells me what I need to do. My productivity doubled."
He Lu — Service coordinator, 1.5 years experience
Beyond CSS
The design system built for this project — the modular components, the information hierarchy patterns, the risk-flagging architecture — was adopted by three other internal platforms at Midea, covering supply chain, sales, and HR.
What started as a service platform redesign became the foundation for how Midea builds internal tools.
Zooming Out
What I actually learned
01
User research isn't about finding problems. It's about finding the right problems.
With 12 PMs and 6 business departments involved, alignment meetings could easily become a bottleneck. I built six rapid prototypes — from conservative tweaks to radical shifts — and ran co-creation workshops rather than review sessions. The goal wasn't to get sign-off. It was to find the edges: technical constraints, business non-negotiables, and tolerance for change. Those workshops gave us a shared map.
02
In complex organizations, design is a negotiation.
Every decision I made touched multiple stakeholders with legitimate, sometimes conflicting, concerns. The co-creation workshops, the incremental rollout, the shadow modules — these weren't just UX tactics. They were ways to build trust while shipping real change.
03
Systems thinking is the only way to do systems work.
The 12 modular process nodes, the three-layer detail architecture, the Glance/Dive/Focus framework — none of these were invented for a single screen. They were built to scale. That's why the design didn't stop at customer service.
