AI-FLUENT PRODUCT DESIGN · VECTORSTICKER · 2026
Designing With AI, Owning the Judgment
AI let me ship a reporting dashboard, new discovery sections and a new taxonomy in 2 months. The harder job was deciding what to trust. All on a live site, with zero downtime.
Role
Senior Design Consultant
Timeline
2 months (Aug to Oct 2026)
Tools
Claude Code, ChatGPT, Google AI Studio, Google Stitch
Focus
Data, Discovery, Information Architecture
2 months
Timeline
5 + 5
Dashboard panels + GA4 views
4
New discovery sections
0
Main site downtime
The short version: 3 calls that mattered
1
I checked real data before building on it.
2
I cut what the data could not support.
3
I set six rules before any code touched the live server.
Chapter 01 · The Ask
One Server. One Live Site. No Room for Mistakes.
It started with a simple ask. The Marketing VP needed a clear view of how the platform was performing, and there wasn't one.
VectorSticker.com is a free platform for vector T-shirt designs, stickers and wallpapers, with 1,788 published items. A previous developer built it on a Laravel marketplace script. Early on, the code moved into a GitHub repository I manage, so the product no longer depended on one person.
The data was there. Nothing turned it into insight.
Reporting
Views data and Google Analytics existed, but nothing brought them together in a view the Marketing VP could use.
Measurement
Downloads were stored only as one running total per item, so a download trend was impossible to calculate.
Discovery
Main categories described product formats, not content. Popular designs kept getting attention, while good designs in quieter sub-categories stayed unseen.
Before: data in pieces
nothing connected them
Views data
Google Analytics
Downloads
One running total per item
Main categories
Product formats, not content
No clear view of how the platform was performing
The catch
Whatever I built would run on the same server as the live site.
Chapter 02 · The Temptation
AI Made Everything Fast. That Was the Risk.
Four AI tools supported the work. They removed the wait between an idea and a working product.
But speed has a side effect. Moving fast makes it easy to build on assumptions. So my role was not to build faster. It was deciding what to trust, what to show and what should never ship.
Claude Code
ChatGPT
Google AI Studio
Google Stitch
How every step worked
AI builds. I decide.
Idea
A panel, a section, a fix
AI builds fast
Four AI tools
I check real data
Is it true?
Ship
Cut
Chapter 03 · The First Twist
Three Assumptions. All Wrong.
Early on, three things looked true. None of them were.
What looked true
What real data showed
Active item code: 1
The thumbnail field holds the images
A guessed image URL works
Each one was caught the same way: by checking real data before building on it.
Then came the bug that wasn't a bug
A login field looked prefilled, like a code bug. A DOM check showed it was browser autofill. Not a code bug.
Another time, a failed database login was traced to a host mismatch between socket and TCP connections. Find the cause first. Change code second.
Evidence first, fixes second.
Two debugging paths
Login field looked prefilled
DOM check
Browser autofill
Not a code bug
Failed database login
Traced the connection
Host mismatch: socket vs TCP
Cause found before changing code
Chapter 04 · The Turning Point
Six Rules Before Any Code Shipped
The dashboard and the main site run on the same server, so every boundary mattered. Before any code shipped, I set clear rules.
Separate surface: the dashboard runs on its own subdomain and server configuration, separated from the main site.
Limited access: its database user reads main site data and writes only to its own login table.
Hands off: the admin panel built by the previous developer stays untouched.
Scoped restarts: only the dashboard's own process is ever restarted. Shared services never are.
Checked upgrades: every runtime upgrade is checked first against the main site.
Safe experiments: Google AI Studio runs on dummy data, and changes go through a separate Git branch before reaching production.
Every change took the same path
Google AI Studio
Dummy data only
Separate Git branch
Against the main site
Every runtime upgrade
Production
Restart the dashboard's own process only
These rules held for the full 2 months. The main site never had a minute of downtime.
Chapter 05 · Building, and Refusing to Build
The Best Panel Was the One I Cut
The easiest chart to build was a downloads trend. I didn’t build it, because no daily download history exists.
New-user and top-creator panels were cut too. Their data was not meaningful. Showing less, but only what is true, was the point.
Every panel faced two questions
less, but only what is true
Is there real data behind it?
Yes
Is the data meaningful?
Yes
5 panels
Plus 5 GA4 views and 2 full report pages
No
No
Downloads trend
No daily download history
New users, Top creators
Data not meaningful
reports.vectorsticker.com went live with its own subdomain, SSL and secure login. It has five panels.
Google Analytics 4 adds five more views: traffic trend, traffic sources, top pages, devices and countries.
As the project grew, the backend was rebuilt from PHP to Node.js, with a React frontend.
Total Views
Total Downloads (cumulative)
Views Trend with 7, 30 and 90 day tabs and hover details
Top Items with thumbnails, links to live items and a full list page
Category Performance with main and sub-category tabs and a full list page
User flow: the Marketing VP's path
Secure login
reports.vectorsticker.com
5 panels
Total Views, Total Downloads
Views Trend
7, 30, 90 day tabs
Hover details
Top Items
With thumbnails
Full list page
Live item
Category Performance
Main and sub tabs
Full list page
Sources
Main site databaseRead only
Google Analytics 45 views
reports.vectorsticker.com
Node.js backendRebuilt from PHP
React frontend5 panels, secure login
Audience
Marketing VPA clear view of performance
less, but only what is true
Views trend with hover details
Top items with live links
Main and sub-category performance
Google Analytics 4 views
Top Items full list page
Category Performance full list page
Chapter 06 · Noticing Who Was Left Out
Popular Designs Kept Winning. Good Ones Stayed Invisible.
With performance finally readable, I looked at the catalog itself.
“
"My own analysis of the catalog surfaced one clear weak point: popular designs kept winning attention, while strong designs in quieter sub-categories rarely got seen."
“
So I gave quieter content a way in. Four homepage sections now share the space:
Top Downloads
the most downloaded designs
New Items
the latest additions
Hidden Gems
4 random designs from underexposed sub-categories
Worth a Look
4 random designs from mid-tier sub-categories
More attention
Less attention
Top Downloads
the most downloaded designs
Worth a Look
mid-tier sub-categories
Hidden Gems
underexposed sub-categories
User flow: a visitor's path to quieter content
reshuffles on every refresh
Homepage
Underexposed sub-categories
Worth a Look
Mid-tier sub-categories
"View more"
Sorted by least viewed
Item card
Shows its sub-category
Similar designs
Hidden Gems and Worth a Look reshuffle on every refresh. Their "View more" links sort by least viewed, giving quieter content a real path to discovery.
Top Downloads and New Items
"View more" sorted by least viewed
One small change
Item cards now show the specific sub-category instead of the broad main category, falling back to the main category when none is set, so users can jump straight to similar designs.
Chapter 07 · Rethinking How People Look
Organized by What Designs Show, Not What Format They Are
The old categories answered one question: what format is it? The better question was: what does it show?
I replaced the four format categories with six content categories, holding 50 sub-categories.
Before
Free T-Shirt Designs
Free Stickers
HD Wallpaper
Seasonal
After
Animals & Nature
Characters
Festivals & Seasonal
Memes
Objects & Symbols
Traditional
50 sub-categories
How every design is classified
A design
What is its dominant visual element?
1 of 6
One category per design
Within its category
50 in total
Two rules keep it consistent
One category per design, classified by its dominant visual element.
Mythology got the same thinking. It moved from broad regions to 8 specific cultures, because people search by culture, not by region.
Broad regions
Greek/Roman
Japanese
Egyptian
Chinese
African
Aztec-Maya
Southeast Asian
Norse
by culture, not by region
Six content categories
Culture-specific mythology categories
Chapter 08 · The Ending
Two Months Later: Zero Downtime
The live site stayed safe the whole way. Here is what shipped around it.
Reporting
1
live reporting dashboard
5
panels and 2 full report pages
5
Google Analytics 4 views
Discovery
4
new homepage sections
Sub-category labels
on item cards
Information architecture
6
content categories
50
sub-categories
8
culture-specific mythology categories
Reliability
Zero
downtime on the main site
Chapter 08 · The Ending
From Reports to Alerts
A dashboard you have to check is still passive. Next, it will speak up.
It will flag when views move more than 20%, social traffic more than 25%, or category downloads more than 20%, with a volume filter to ignore noise. A daily snapshot of item data comes first, creating the download history that never existed. On the content side, art styles will move into tags, and each mythology category will get search-friendly titles and descriptions.
Planned
concept sketch
Views
Social traffic
Category downloads
Planned alert flow
Planned
Daily snapshot
Item data
Compare change
Views, social traffic, category downloads
Volume filter
Ignore noise
Over threshold?
20%, 25%, 20%
Alert
The dashboard speaks up
The Lesson
AI Sped Up Execution. Judgment Shaped the Product.
AI removed the wait between an idea and a working product. It did not remove the need to check, cut and protect.
Looking back, the most important calls were often about what not to build: a trend without data behind it, a panel without meaning, or a change that could put the live site at risk. To me, AI fluency is not about generating code or screens quickly. It is knowing what to ask, what to verify and when the evidence is not strong enough to ship.
“
AI made me faster.
Verification made me trustworthy.
That’s my journey with this project up to this point.
