DevFest
Gaborone
2026
Google
Developer
Groups
01

Un-AI your UI

Building software got faster.
Deciding what's worth building didn't.

Hi, I'm

Rudisang
Morake

Software engineer
at Leverage Point

Rudisang Morake
01

Even before AI, UI was an afterthought

“Yeah, the UI is ugly.
But at least it works.”

Get it working first. Make it look good later.

Database, logic and API got the planning. The UI got whatever time was left.

And we got away with it, because for the most part it worked.

Week 12345678
Database
Logic
API
How it lookslater
A table you scroll sideways
A Save button that never says it saved
Every country on earth, no search
But it works

With AI, that habit
now happens at scale

Building an app used to be hard.It took real skill and months of work. Few people could build a better one, so a bad UI survived.
Now anyone can pull the lever.Anyone and their mum can wake up and ship an app by lunch.
The result: UI litter.Everything gets shipped. Nobody sweats the finer details.
DEVELOPERS ONLY
Pull again: same dashboard, slightly different purple

It looks finished, even when
nobody decided anything

Old UIs looked unfinished, so people asked questions.

AI output comes out shiny, so it sails through review.

Underneath, hundreds of small calls nobody actually made.

then

Expense Tracker v0.3 (beta)

DateItemAmount
12/11/2012Groceries350.00edit | delete
13/11/2012Taxi12.5edit | delete
13/11/2012airtime50.00edit | delete

Item: Amount: Submit

Error: amount must be a number!!

“Who designed this? Not ready yet.”
now
Who is Alex? Why these three numbers? What is this chart for? Is there even a Pro plan?
Approved: “looks great!”

“At least it works” doesn't
set you apart anymore

then
It works

Getting it working was the hard part.

now
It works
Someone thought about the person using it

An agent gets most things working in an afternoon. Everyone starts here.

02

You've probably seen this website

Meet DoMorely. I made it up for this talk.

✦ THE FUTURE OF PRODUCTIVITY

Unlock your
full potential.

Your all-in-one platform to streamline workflows, supercharge productivity, and transform the way you work.

⚡ 10x faster workflows
✓ 99.9% uptime
★★★★★ Loved by 10,000+ teams
✧

Seamless integration

Connect everything. Unlock possibilities. Make your workflow work for you.

↗

Powerful insights

Turn information into impact with the tools you need to do more.

⌘

Effortless collaboration

Bring your team together and take productivity to the next level.

So what does it actually do?

A budgeting app?A project tracker?A chatbot for your plants' feelings?

Swap in another name and the page still works.

So the design isn't saying anything about the product.

The familiar AI DNA

Never one thing. A bundle of defaults that keep showing up together.

None of this proves a model made it. It shows someone went with the defaults.

  1. 1A purple gradient, a glow and a giant centred headline
  2. 2A pill that says “THE FUTURE OF” something
  3. 3Floating badges, everything rounded
  4. 4Three identical cards, icons in rounded squares
  5. 5Copy that unlocks, empowers and streamlines
✦ THE FUTURE OF PRODUCTIVITY

Unlock your
full potential.

Your all-in-one platform to streamline workflows, supercharge productivity, and transform the way you work.

⚡ 10x faster workflows
✓ 99.9% uptime
★★★★★ Loved by 10,000+ teams
✧

Seamless integration

Connect everything. Unlock possibilities. Make your workflow work for you.

↗

Powerful insights

Turn information into impact with the tools you need to do more.

⌘

Effortless collaboration

Bring your team together and take productivity to the next level.

03

Does it matter if it looks AI?

Yes and no.

Yes: people skim your page like a CV in a pile

200 CVs on the desk. Each one gets a few seconds.

One says what this person actually did. That one gets read.

Your homepage is your product’s CV.

No: a familiar layout is fine.
Two hundred identical ones aren’t.

A normal CV layout helps. The recruiter knows where to look.

AI can write two hundred of the same CV in an afternoon.

On the web, that's cards, rounded corners and a sensible font.

They worked long before AI. The models learned them from us.

04

The problems that matter more than the purple

The code works. Nothing crashes. No test fails.
The design still lets someone down.

Dark mode can change
what your screen means

Which account is selected?
Warning hard to read
Chart colours picked for white

Give each colour a job, not a hex code:

text selected warning success

Put a word or an icon next to any colour that means something. Check both themes.

NN/G, Dark Mode (2023). WCAG 1.4.1: colour can't be the only signal.

Correct numbers can still
tell the wrong story

bars start at 0%

100%50%0%
94%
96%

bars start at 93%

100%50%0%
94%
96%
Looks three times taller
The real change is 2 points. Nothing tripled.

UK ONS: bar charts start at zero, because the length is the comparison. Made-up values.

Your imaginary user has
a very easy life

The Alex we design for

  • Short name
  • Perfect eyesight
  • New laptop, two free hands
  • Nowhere else to be

The real Alex

  • Phone in one hand, bag in the other
  • Bus in 2 minutes
  • Pale grey text hard to read
  • Tiny icon hard to hit
  • Hover tip phones don't hover
booking form

Last name is required

A page can look right and
answer the wrong question

Asked for a “profile page”

9:41
LM

Lerato Molefe

Coffee, code and sunsets ☀

248posts1.2kfollowers310following

What the student needed

9:41

Hi Lerato

Due next

Essay draftFriday, 17:00
Quiz 3Next Tuesday
Group projectIn 2 weeks

Your courses

Design for the web72% done

Lerato Molefe, profile

Before asking if it looks good, ask what someone should notice first.

NN/G, AI Prototyping in Real Design Contexts (2025).

05

How we can un-AI our UIs

Give the model better direction.
Give it good examples.
Check the decisions it makes.

A wish leaves the model guessing

Ask for an app…
+CanvasEffortLow

Effort

LowMediumHighXHighMax
result
Thinking…Thinking…
Who is it for? What matters most? What should it look like?
“Make no mistakes” is like telling a striker “just don't miss”.

Anything you don't specify, the model fills in with the most common answer it has seen.

A brief gives it decisions to follow

Ask for an app…Design the home screen of a budgeting app for someone checking their money on a phone before making a purchase. Their first question is “What can I safely spend before payday?” Make that amount the clearest thing on the screen, with bills due next and a short explanation of how it was calculated. Use BWP amounts. Keep the existing brand colours and type. Use plain labels, readable contrast, and text alongside warning colours. Show light and dark versions with the same priorities, plus a view with larger text. Before building, tell me what you're assuming about this person and their task.
+Canvas
who it's fortheir questionwhat matters mostwhat not to changewhat to test withsay your assumptions
lightdarklarger text

Follow up on the decisions,
not the looks

1

2

3

Prompts only get you as far
as your design knowledge

You can't ask for a fix to a problem you don't know exists.

A skill brings the know-how of people who've already solved these.

what you know to ask for Brand colour Make it responsive Dark mode What gets noticed first Line length Colour that carries meaning Keyboard focus A crowded table on a phone Touch targets Contrast in dark mode Empty and error states Alt text that means something Loading states
design-skill/
what the skill brings

A skill writes that know-how
down once

frontend-skill/
SKILL.md: how to approach the job References and examples Scripts Steps to check what it built
for example: Taste v2

Turns “have better taste” into things to do

  • Concrete directions for composition, type, colour and imagery
  • Inspects the page and reviews screenshots of what it built
  • References drawn from 247 curated gallery sites

Most model providers ship one too.

Skills have opinions.
That's the help and the risk.

Night
Market

Fridays from 6pm. Food, music, makers.

Get a stall

Marketing page: memorable

Payroll

Run 1 of 3. Review 214 rows.

Approve

Screen used all day: exhausting

A ban on gradients is a house preference, not a law of usability.

Read the skill you install. No general skill knows your users or your brand.

So we wrote two skills of our own

Slopless

helps while we build
  1. Reads what's there: a real design, an untouched starter, or nothing yet
  2. Asks a few plain questions, then picks a direction
  3. Applies house rules for layouts, forms, labels, nav and delete actions
  4. Finds something specific first: payday, not “premium SaaS energy”

Slop Police

checks what came out
  1. Scanner for the mechanical tells
  2. Checklist for what a scanner can't see
  3. Screenshots it actually looks at
Enforce fixes new workAudit suggests, changes nothing until you pick

The scanner catches about half. The review asks: does it make sense, and does it still look good?

1 / 3

Built without the skillsSame budgeting app brief. AI defaults everywhere.

Retrofitted with the skillsSame site, same brand colour and font, every tell replaced.

Built from scratch with the skillsIts own direction, built around payday.

What we found

Without the skills
51
scanner findings
Retrofitted
2
both kept on purpose
From scratch
0
findings

Scanner findings on a landing page and a settings page. Not a usability study.

A blank page also scores zero.

Our first try scored 0 and looked thin. We'd taught the agent what to remove, not what to put back.

Every tell was doing a job. Replace the job:

Feature-card grid→A real preview of the product
Glow→Scale, composition, a proper block of colour

Show the agent your design.
Don't describe it.

style-guide.md

# Style guide

Use clear hierarchy, comfortable spacing, and consistent buttons.

clear how?comfortable how?consistent with what?
style-guide.html

Safe to spend

Body text at the size people will read it.

Save changesCancelFocusedDisabled
P 9,850
P 0

Enter your rent, or 0 if you don't pay any.

Markdown for the why HTML for the how

None of these is the whole answer.
Use the combination.

a specific prompt

The person, the task, the priority, the constraints

a skill that fits

Repeatable decisions, references and a review process

an HTML guide + written rules

Examples to inspect, reasons to follow

human review

Real content, in the conditions people use it

Start with one important screen

  1. Decide what someone needs to understand first
  2. Give the agent that context
  3. Use a skill that fits
  4. Point it at a real example
  5. Look at the result where people will use it

Make a UI where you can tell someone thought about the person using it.

Thank you

Rudisang Morake
Software engineer at Leverage Point

Go to slide Enter
00:00/ 30:00
end

next slide