Senior Product Designer
Research · UX · UI · Front-end build
Background
Profitable Meta acquisition runs on creative volume.
Roundforest wins on Meta by testing many ad creatives every day. I led UX and UI for CopyCat, the internal AI platform that turns a product task into a landing-page article and a set of ad creatives.
Problem
A faster process that became harder to run.
A no-code system had already lifted output by 30%. But it lacked context and didn't fit the team's workflow, so they stitched together six tools around it. Production got faster and much more complex.
70–100today
~130breakeven
500+profitable scale
BeforeGoogle SheetsAsanaGeminiIdeogramKueez CMSCopyCat (old)
AfterCopyCat · one end-to-end workflow
If we fix efficiency, we can at least double output.
Research
Two weeks inside the Meta acquisition team.
I joined the team for two weeks to experience the workflow from the inside: the real pain points, the constraints, and the gaps between their tools and what they actually needed.
Mapping
Mapping the workflow showed where it broke.
I mapped everything from ideation to campaign launch. It surfaced redundant steps, tool fragmentation, and clear room to consolidate and automate.
Ideation→Article→Creatives→Campaign→Testing
Unify
Replace six disconnected steps with one continuous flow.
Automate
Data handling, layout and asset generation, with no manual work.
Review
Keep people in charge where quality and cost are at stake.
Decision
I chose a guided wizard over one big screen.
The biggest call was how to structure creation. Keeping the familiar layout was safe. I chose to redesign it, knowing it carried a real risk.
Option A
Improve the existing screen
Every tool and setting always on display, with no enforced order.
Keeps the clutter, and steps stay easy to skip.
Option B Chosen
A guided wizard
A fixed sequence where each step unlocks the next.
Risk: the team relearns a tool they use daily.
The five steps
Setup → Templates → Generate → Review & select → Download
1 / 5
The team changes often
A tool that teaches its own process beats familiarity.
Order prevents errors
No more skipped or out-of-order steps.
Short learning, long payoff
The structure keeps paying off as volume grows.
To get the PM and engineer on board, I built the new front end myself on the existing system, so the redesign added no backend work.
Solution
One task list drives the whole flow.
Every day starts from the task list. Each task carries the data needed to generate its article and its creatives, so nobody re-types anything between steps.
Generate article→Generate creatives→Review & approve
Single source of truth
Task rows hold all metadata, so creation needs no manual entry.
Intentional review gates
Generation costs money, so a person approves before anything moves on.
Clear status
Finished tasks turn green before handoff to PPC.
Dashboard tab
Throughput and bottlenecks by owner and category, with no spreadsheet export.
The whole flow in 30 seconds
From a task row to an article, through the five creative steps, and back to the list.
Deep dive
Dashboard
Completed and open tasks by owner, section and category. The team sees throughput and bottlenecks without exporting to a spreadsheet.
Article generator
AI drafts the landing-page article from the task's page title. Every field shows where it came from and can be regenerated on its own.
Templates
Templates matched to the product category come pre-selected. Uncheck what you don't want, then generate everything in one run.
Review & select
Each creative is tagged Good, Needs review, or with a specific issue. Auto-pick the best, confirm the checklist, and only then download.
AI review
AI drafts. People decide.
CopyCat uses GPT, Gemini and Ideogram for article text, banners and ads. Quality varies and every generation costs money, so the interface treats AI output as a first draft.
Label the source
Fields are tagged "auto-filled from task" or "AI generated".
Regenerate one field
Fix a weak line without losing what works.
Options, not one answer
Three variants turn reviewing into choosing.
Edit, swap or upload
Everything stays editable, including the banner.
A person saves
Only "Save Article" moves the task forward.
Designing for when AI gets it wrong
States for the moments that break trust: waiting, failure, and output that looks fine but isn't.
RegeneratingLoading
Previous version kept · Restore
Only that field loads, and the old version is one click away.
Generation failedError
The image service didn't return a banner. Your other edits are saved.
Try againUpload
The error stays in its field and always offers a manual way forward.
Automated checksNext phase
The guaranteed fix for a cluttered garage
"Guaranteed" may break Meta ad policy
Flag likely policy issues before a person reviews.
These states were designed for this case study to extend the shipped flow. Automated checks are planned for a later phase.
Outcome
Same team. Seven times the output.
Within days of launch, daily output passed 700 creatives, clearing the 500+ needed for profitable scale.
700+
creatives a day, up from 70–100.
75%
less time per article and creative.
6 → 1
tools consolidated into one platform.
–36
manual clicks removed through automation.
Production time per article and creative, before vs. after launch. Daily output is the average over the first week.
Learned
Separate the interface from the system. Test the bet earlier.
Make big changes cheap for the team. The wizard was a hard sell until I showed the features and backend would stay the same, and then built the front end myself to keep that promise.
Validate before committing. With one month, the wizard was proven mostly by real use after launch. Next time I'd explore more iterations and run usability tests with daily users first.
Try it
Generate an article, then build the creatives.
The same task flow that shipped. Generate an article to unlock the creative builder.
The interactive prototype is built for larger screens. Open this page on a computer to click through it.