How I work with AI, in plain language

What this is about

Two things that always come up.

Both are fair questions. That is why they go first, with an answer instead of a shrug.

AI takes jobs away.

The work is changing, true. Denying that would be odd, because I am the designer and the developer who used to do this job. My own trade shifted under me.

What disappeared was mostly the waiting: the weeks between a good idea and something that works. What did not disappear are the decisions. Those got heavier, because more comes out faster, and somebody still has to decide what ships and what does not.

I could do that too. One prompt and done.

One prompt gets you something that looks good. That is not the same as something that stays up with a thousand people on it, with money running through it, or when it needs changing a year from now.

The difference is not in the first version. It is in the hundred decisions after it, and in knowing which of the four suggestions to refuse. No prompt helps with that. Experience does.

From idea to live

The prompt is step four.

Five steps, and who does what in each. You can see straight away where the machine is strong and where it has no business being.

Bureau met een half afgebouwde interface op het scherm en handgeschetste wireframes in een notitieboek

N° 01

The idea

Why does this thing exist, who is it for, and what must it never become. No screen and no line of code involved yet.

nothing

everything

N° 02

The structure

What gets stored, how it connects, and what we deliberately leave out. The dullest step by far, and by far the most expensive to fix later.

proposes a setup

decide what goes in

N° 03

The look

How it looks and how it feels. A model gives you ten variations in seconds. Nine of them are tidy, correct, and exactly what everyone already has.

delivers variations

bin nine of them

N° 04

Building

Only here does the machine really get to work, writing most of the code. Not in one instruction, but in hundreds of small ones, file by file, with me in between.

writes most of it

steer, file by file

N° 05

Checking and running

Does it work, is it safe, and does it hold up when somebody pulls at it. After that it has to keep running for months on a server that is mine.

reviews along

am accountable

Under the hood

What I watch for.

Six things a model does not get right on its own. None of these mistakes show up if you only look at the screen.

N° 01

The database

A model proposes a setup that works fine with ten users and grinds to a halt at ten thousand. I look at what is going into it: which data, how it connects, and where searching has to stay fast.

If nobody checksYou only notice once you are stuck with it, and by then rebuilding costs more than starting over.

N° 02

Where your data stays

BankBird reads bank statements. Data like that does not go to some other party, and is not used to train models. I decide per field what gets kept and what goes straight out.

If nobody checksYour data ends up with services you never chose, and you will not get it back out.

N° 03

Security

Code that works is not the same as code that is safe. I go through who may see what, what happens with odd input, and whether anyone can reach data that is not theirs.

If nobody checksNobody notices until it goes wrong, and then it goes wrong all at once.

N° 04

Back-ups that work

A daily back-up is the easy part. The real work is checking now and then that you can actually restore from it, because a back-up you have never restored is an assumption.

If nobody checksYou find out on the worst day of the year that the file was empty.

N° 05

What you pull in

A model happily solves a problem by pulling in ready-made parts from other people. Every part is something you did not write and still have to trust. I keep that list short.

If nobody checksYour project hangs on dozens of scraps of code from strangers, and one of them will eventually be a leak.

N° 06

Speed and weight

This site once had a list of small images 108 pixels wide that quietly loaded 10 MB of originals. Everything worked, everything looked good, and on a phone it crawled.

If nobody checksIt runs beautifully on the maker's computer and nowhere else.

The same job

What this would have taken with people.

Take one product from the range, say BankBird. These are the roles normally at the table for that, and the lead time that comes with them.

Lege vergaderzaal met vijf stoelen, laptops en uitgeprinte wireframes op tafel

Role With a team With me

Concept and structure With a team1 to 2 weeks With meMichael

Design and interface With a team2 to 3 weeks With meMichael + AI

Front-end build With a team3 to 4 weeks With meMichael + AI

Back-end and database With a team4 to 6 weeks With meMichael + AI

Testing and bugs With a team1 to 2 weeks With meMichael + AI

Project management With a teamongoing With menot needed

All together 3 to 4 months4 to 6 people Weeks1 maker

An indication, not a quote: every project runs differently. And it is not the whole story, because there is something that does not get shorter.

What does not get shorter are the decisions. Twenty years of designing and building sites is exactly why this works: without those twenty years you would not know what to ask for, and more importantly, what to refuse.

No key held hostage

And if AI disappears tomorrow?

PHP Laravel JavaScript HTML CSS MySQL

Then nothing about the code changes. Everything I develop sits in languages and frameworks that have been around for decades. No invented language, no trick that only works for me, and no black box nobody can look into.

A developer you hire next year opens the folder and sees a project they recognise. They do not need to use AI themselves to carry on with it. That is exactly the point: you are not tied to me, and not to a model.

All of it tooling every web developer knows. Here AI is the accelerator, not the foundation.

The limit

What AI does not do.

Fair is fair. These are the places where the model leaves you standing, and where the work stays plain old work.

Scherm met een voorstel dat is afgekeurd en weer overnieuw gaat

01

It does not know what you mean

Only what you type. Half explained means half built, and the model will not say a word about it.

02

It invents with confidence

An answer that makes no sense sounds exactly as certain as one that is right. You have to know yourself which of the two you are reading.

03

It has no taste

It neatly averages everything that already exists. That is precisely why so many AI sites look alike, and why most of it ends up in my bin.

04

It carries no responsibility

When something falls over on a Sunday evening, no model gets up to fix it. I do, with my name on it.

05

It does not remember your project

Why something was built a certain way three years ago is nowhere inside a model. It is in the maker's head, or it is gone.

AI does the work. I make the calls.