
THE SOURCE · 2026
One approved record for every word.
An internal tool I designed and built for a UAE Islamic bank. It carries interface text from Figma to code.
The bank is not named, and the product’s name has been changed, for confidentiality.
80%
less manual QA on UX copy, in a one-week demo
60%
shorter approval time, in a one-week demo
2,900
approved texts in English and Arabic (approx.)
98 → 26
addresses the bank’s gateway team registers by hand
MY ROLE
Concept, research, design and build.
Senior Product Designer on a product with two parts: a Figma plugin for designers and a web application for everyone else. It ran from March to November 2026, through five versions.
Who it serves: designers, product owners, department reviewers, engineers, viewers and administrators. Business owner: the bank’s digital transformation team.
The business goal
Replace a chain that ran by hand and left no record with one recorded path for every text, in English and Arabic.
Manual QA on UX copy
−80%
Approval time
−60%
From a one-week demo. The track is the old process, and the red bar is what remains with The Source.
The bet I started with: build the first version quickly on outside cloud services, and prove the idea.
Result: it proved the idea and could not be used. A security and architecture review found that every part of it broke the bank’s rule: all data stays inside the bank.
EVIDENCE
Seven checks, each tied to a decision.
Security and architecture review
Review of the first version
v1
Every part broke the data rule. I rebuilt it on components the bank already hosts.
Gateway address count
Every address the web application calls
98 → 26
I restructured how the application talks to its server, without removing a function.
Structured test plan
Run on the fifth version, with a register of edge cases
6 roles
The full path from design to code is covered for every role.
Security self-assessment
Sign-in, access control and data handling
0
No weaknesses found.
Adoption documents
Design, tests, operations, security, business case, handover
16
A bank adopts a system through documents, so I treated its reviewers as users.
Approved library
Records in English and Arabic
≈ 2,900
The record every new design is checked against.
One-week demo
Manual QA on UX copy down 80%, approval time down 60%
1 week
Indicative. The pilot will set a measured baseline.
Every recording on this page uses made-up demo data, never real bank wording.
01 · THE PROBLEM
Nobody could point to one place and say: this is the approved text.
Every customer-facing word in the bank’s apps has to be written, reviewed by several departments, produced in two languages, and built exactly as approved. That chain was manual and unrecorded.
Wording was agreed in conversations, then copied by hand from the design into the apps.
None of this was measured.
That became a design requirement in itself: the product had to make the process visible enough to measure.
What was happening, and what it cost
No single record of what was agreed
Time spent reconciling versions, and decisions made twice
Reviews stalled without anyone noticing
Product owners chasing approvals, and late launches
English and Arabic drifting apart
Text shipping with one language missing or misaligned
Text typed by hand into the apps
Engineering rework, and wording in production that differed from what was approved
The same wording rewritten for each project
Departments reviewing text they had already approved
Approval evidence scattered
Slow answers to audit and regulatory questions
02 · WHO IT IS FOR
Nobody should have to learn a new place to work.
Six kinds of people touch a piece of text before a customer reads it, and each one needed something different from the product.
Designers stay in Figma. Reviewers get one page with the text, a screenshot of where it appears, and three choices.
Six roles, and what each needs
Designer, in Figma
To check text and send it for review without leaving the design
Product owner, on the web
To choose who reviews, see where each text is stuck, and publish when it is cleared
Department reviewer, on the web
To see the text in context and approve, edit or return it quickly
Engineer, on the web
To download approved text in the format their platform uses
Viewer, on the web
To follow a project without being able to change it
Administrator, on the web
To manage people, departments and business areas
03 · THE PATH OF A TEXT
One path, four stops.
The Source gives every text one path with four stops: checked in the design, reviewed by each department in order, published to one library, and exported to code.
The dashed arrow is what makes the library pay back: wording approved once is recognised in the next design and skips review.
04 · CAPTURE FROM THE DESIGN
Text that is already approved does not go round again.
The plugin scans the selected frames and tests each text against the bank’s writing rules and the approved library.
Everything else is sent for review with a screenshot of the frame and a link back to it. The reviewer sees the words where the customer will see them.

The plugin scans a frame: five texts, one flagged and one already approved. The designer changes “Click” to “Tap”, scans again, and the issue clears.
05 · REVIEW IN ORDER
Each department takes its turn.
The product owner picks a workflow, for example Marketing, then Customer Experience, then Shariaa.
A department can also approve a whole project in one action when every text is fine.
Three choices for a reviewer
Approve
The text moves on to the next department.
Edit
Only the language their department owns.
Send back
To a named earlier step, with a reason.

A Customer Experience reviewer opens a text and approves it with a note. The text moves on to Compliance.

Sending back is specific: the Compliance reviewer returns the text to Marketing, with a reason.
06 · ONE APPROVED LIBRARY
Nearly the same wording is a decision, not an error.
When the last department has approved, the product owner publishes the text to the library: the single record of approved wording in both languages, with the history of every change.
Two checks run at this moment. Wording that nearly matches an existing record is shown word by word against it. Text with a language missing is held.
When new text almost matches a record
Add as variant
Same message, accepted alternate wording. It is stored on the existing record.
Add as new record
A different message. A separate approved text is created.
Cancel
Decide later, or fix the wording to match exactly.

A near match at final approval. The one added word is highlighted, and the product owner keeps it as a variant.
07 · HAND OVER TO ENGINEERS
Nothing is typed by hand.
Engineers download approved text in the format each platform uses: Android, iOS and web.
English and Arabic come as two files sharing the same keys. Engineers can ask for only what changed since the last download.

Export for engineers: pick Android, iOS or web, and optionally only what changed since last time.
08 · WORKING INSIDE A BANK
The hardest design work was not on a screen.
It was fitting the product to the rules a bank runs by.
I rebuilt the first version on components the bank already hosts. One feature in the plugin relied on an outside AI service. I removed it instead of sending bank text outside.
Every address the web application calls must be registered by hand by the bank’s gateway team, and a proof of concept has no budget for that. The first count was 98. I restructured how the application talks to its server and brought it to 26, without removing a single function.
The bank’s reviewers are users too
A bank adopts a system through documents, so I produced the set it needs: 16 in all.
Design
Test plans and results
Operating procedure
Security self-assessment
Business case
Handover checklist
26
addresses to register by hand, down from 98. That is 72 fewer for the bank's gateway team.
One question is still open. Security policy does not allow a tool inside Figma to connect into the bank’s network.
The answer I am designing is a plugin that works offline and exchanges files with the web application.
DESIGN DECISIONS
Five decisions shaped the product more than any screen did.
A review is a snapshot, not a live template
Assigning a workflow copies it into the project. Changing the template later cannot rewrite a review already under way, and an approved step cannot be reordered or removed.
Stop 2, review
Sending back is specific
A reviewer returns text to a named earlier step with a reason. The steps in between review it again, so nothing slips through on an old approval.
Stop 2, review
Nearly the same wording is a decision, not an error
The product owner sees the two compared word by word, then keeps it as a variant, saves it as a new record, or cancels.
Stop 3, library
A missing language stops the line
Text cannot be published with its Arabic or English missing unless someone chooses to override, and the override is recorded.
Stop 3, library
What the screen promises, the system enforces
Hiding a button is not a rule. Publishing is impossible until every department has approved, and the name on each approval is set by the system, not by the sender.
Every stop
OUTCOME AND NEXT
Five versions, built and tested. A pilot is next.
The fifth version covers the full path from design to code for all six roles. A structured test plan has been run and its results recorded. A security self-assessment found no weaknesses in sign-in, access control or data handling. In a one-week demo, manual QA on UX copy fell by 80% and approval time by 60%.
What I would verify next
A three-week pilot on retail mobile and internet banking. The old process was never measured, so the pilot first sets a baseline, then tests six expected benefits:
Faster approval of content, from first draft to published.
Less effort spent chasing, reconciling and retyping.
Approval evidence that can be produced on request.
More text released complete in both languages.
Fewer differences between approved wording and what ships.
More reuse of approved wording across projects.
What I learned
Constraints are users too.
The security reviewer and the gateway team never open the product, yet their needs shaped it as much as any designer’s or reviewer’s did.
A prototype proves the idea. The rebuild proves it can live.
The first version won the argument that the product was worth having. Almost none of it survived contact with the bank’s rules, and that was the right outcome.
Building it myself changed how I design.
I built The Source with an AI coding assistant, directing it in plain language. Questions a mock-up never raises became unavoidable: what happens when two people edit the same text, or when a workflow changes in the middle of a review.
Say what is not yet proven.
The business case asks for a pilot, not for belief. Naming the missing baseline made the proposal easier to trust.