For everyone building with AI
The bottleneck moved.
Product development, now that building with AI is the cheapest part.
Where I am coming from
Two decades designing for a billion people. Now I build with the same tools you do.
LinkedIn, Microsoft, eBay, PayPal, Wells Fargo, Trust & Will, Secure Quantum Services, and more. The tools are probably the same ones you have open right now. What they do not give you is the process. I spent twenty years inside one, and this talk is what survived the move from big teams to just me.
US patents on UX architecture, AI systems, and mobile
monthly active users on products I designed and launched
cumulative revenue influenced across those companies
design awards, local and international
How it used to work
The build was the expensive part.
Every gate in the old process existed because engineering time was scarce. You had to earn the right to build. Research, definition, design, review. The gates felt like friction. They were where the thinking lived.
What changed
A sprint became an evening.
So the gates fell. Nobody removed them on purpose. They stopped being required to reach code, and the thinking inside them went with them. We kept the speed and dropped the judgment.
The reframe
Building is cheap now.
Deciding is not.
Speed is the entry fee. What follows is the judgment the tool does not have, and how to learn it faster than the tool can build.
The whole lifecycle
Six stages. AI made one of them cheap and easy.
Discover
Imagine
Evaluate
Research
Plan
Define
Vision
Strategy
Resources
Alignment
Design
Wireframes
Concepts
Revisions
Prototype
Develop
Execute
Frameworks
Program
QA
Deploy
Ramp
QA
Refinement
Launch
Delight
Give users a reason to love their experience, so they come back and tell people
YoursThat is where most builders start. The other five are where the product gets decided, and they are the rest of this talk.
What the tool actually gives you
The AI hands you the median app.
You know the screen. Sidebar left. Cards in a grid. A dashboard you didn’t ask for. That is not real design. It is an average, the mean of everything the AI model was trained on, and it happens to look finished. Looking done is the trap.
- A point of view
- An empty state
- An error path
- A first minute
- A hierarchy
- A brand
- Your name on it
The gap most builders never see
Every screen has five states. The tool builds one.
Nothing yet. The first thing a new user ever sees. Tell them what goes here and give them the one action.
Under two seconds, show shape. Over two, say what is happening.
Three items, not thirty. Missing fields. Half a profile. Does the layout still make sense?
It failed. What happened, what they lost, what to do next. In their words, not the stack trace.
Perfect data, perfect length, perfect light. The only one the demo needs, so the only one you got.
Your users live in the other four. Ask the tool for them by name, one screen at a time, and check each one yourself.
Working through a user experience
A user experience is a path, not a pile of screens.
Write the one job a stranger came to do. Walk it yourself with a real goal and a timer. Count the taps. Mark every place you hesitated. That path is the product. Everything off it is a distraction you are paying to maintain.
Sign-up first is the tool's default, not a decision. Show something useful before you ask for anything. If a stranger cannot reach value in a minute, no amount of features will keep them.
Time to value is the only onboarding metric that matters. Measure it with a stopwatch before you measure it with analytics.
Hierarchy
One thing per screen.
The tool gives everything equal weight because it does not know what matters. You do. Squint at the screen. If you cannot tell what to do first, neither can they.
- 01
Navigation is the path, not the org chart.
Day one nav has three items, maybe four. The eight-item sidebar is the tool guessing at your roadmap.
- 02
Cut anything you cannot explain in one line.
If a feature needs a paragraph, it needs a decision, not a tooltip.
- 03
Default to less. You can always add.
Nobody has ever churned because an app was too clear.
Brand
Brand is the promise the product keeps, screen after screen.
Not a logo. Not a palette. One sentence about who it is for and what it does for them, one voice, one visual system, repeated until people can recognize you with the name covered.
Write the brand down where the AI can read it: a short file in the repo with your one sentence, your voice rules, your banned words, and your color and type tokens. Every session starts on brand instead of on default.
Your nouns are your brand. Decide what the core thing is called before the tool calls it "item."
Copy
The words are the interface.
Oops! Something went wrong. Please try again later.
Nobody knows what happened, what they lost, or what to do. Every app in the world says this, so it reads as no one's voice.
We could not save your changes. Your draft is still here. Try again, or copy it somewhere safe first.
What happened, what they still have, what to do next. In their words. That is a brand decision hiding inside an error message.
Trust is decided in the small text: errors, confirmations, empty states, the privacy policy. When a payments provider reviewed one of my products, the marketing passed clean. The line that nearly sank approval was the first sentence of the privacy policy.
Talking to the tool
Two prompts per feature: one to build it, one to prove it.
Add a settings page with the usual stuff and make it look nice.
Four decisions handed to a machine that will happily make all of them, differently, every time.
Create /settings. Three fields: display name (text, required), email (read only), timezone (select, default America/Los_Angeles). Save writes to the profiles table. Do not change navigation. Do not add fields I did not list.
Narrow. Surgical. Verbatim. Then a second prompt: show me the route, the table, the state after save, and the error path.
Whatever you leave out of a prompt, the tool fills in with average training knowledge. Decide first. Then prompt. Then make it prove it did what you asked it to do, reliably.
Learning through it
Ship. Observe users. Make informed changes.
Instrument before you ship: one metric that means the job got done, not page views. Then watch a real person use it, in silence, without helping, for ten minutes. It is excruciating, and it beats a week of guessing.
Put together
A week that holds.
Read the plan, the one that lives in a file, not in a chat. Pick one feature. Write its build prompt and its verification prompt before touching a tool.
Build. Verify. Draw the other four states. Record what broke. Mark it shipped in the plan, or blocked and say why.
Walk the path with a timer. Fix the hesitations. Anything clever that shows up goes on the parked list, not in the build.
Put it in front of one stranger for ten minutes. Say nothing. Write down where they got stuck.
Write what changed, what broke, what you learned, what is next. Your future self reads it Monday. So does your AI.
Take this one home
Before you accept a screen, ask.
- What is the one thing this screen is for?
- What does it look like with nothing in it?
- What happens when it fails, and what do they lose?
- Can a stranger reach something useful in sixty seconds?
- Whose words are these? Would I say them out loud?
- Does it look like every other app, or like mine?
- Is the navigation the path, or the org chart?
- Can I read it, tab through it, and use it one-handed?
- What did the tool decide that I never decided?
- What did I learn since Monday, and did the plan change?
Take these with you
Three files I use every day. Now yours.
You just built [screen] at [route]. Before I accept it, produce the four states you did not build: empty, loading, partial, error. Change only what each state requires. Then list every place a real user could see something other than the ideal state that you have not handled. Do not add features.
Run it after every build prompt. The list at the end is the tool confessing what it skipped.
Audit [screen] against ten questions. Answer yes or no with one sentence of evidence. Do not fix anything yet. What is the one thing? What does it look like empty? What happens when it fails? Whose words are these? What did you decide that I never asked for?
The full ten are on the previous slide. Question nine is the method.
Read this file before any change to the interface, copy, or styling. One sentence about who it is for. The product's nouns. Voice rules and banned words. Type, color, spacing, radius. Navigation, in order. Every screen has five states. Do not add what the plan did not name.
Lives in the repo. Every session starts on brand instead of on default.
Scan, or go to product.eleven11.ai/five-states
Enter your email and gain immediate access to all files.
The whole talk in one line
Anyone can build now. Deciding what deserves to exist is still your job.
Thank you. Questions welcome.