Hi ,

I broke one of the original 1935 panes while restoring a window in

my house.

Before I started I bought a 20 page guide from a building

conservator with over 40 years of experience. Every piece of his

advice pointed the same way. Save the original material. Old glass

has a patina you cannot buy back.

When the pane broke I called a glazier. He told me nobody tries to

rescue an old pane. The labour costs more than breaking it on

purpose and cutting new glass to size.

Two experts. Opposite advice. Both right.

They were optimising for different things, and neither of them

could tell me what I actually cared about. That turned out to be

the same problem I have with AI coding tools most weeks, and it is

what this week's post is about.

Read it here:

https://empatheticdeveloper.wordpress.com/2026/09/13/the-glazier-was-right-too/

Three things I have written before that sit underneath this one.

1. Fluency tells you nothing about correctness.

In "Under the Hood" I went through what these tools actually are:

pattern completion over an enormous amount of text. That matters

here, because a tool that quietly renames things into ASCII does

not sound uncertain while it does it. It sounds the same as when

it is right. The wrong artifact arrives with full confidence and

normal formatting.

https://empatheticdeveloper.wordpress.com/2026/03/23/under-the-hood-what-your-ai-coding-assistant-actually-is/

2. Reading the terrain is the job.

"The Code Was Always the Door" called the senior role shepherding:

read the terrain, choose the path, watch for predators. A default

that does not hold for your domain is terrain. It is in your

codebase, not in any training data, and nobody outside your team

is going to notice it for you.

https://empatheticdeveloper.wordpress.com/2026/04/25/ai-shepherd-senior-developer/

3. Where to look first.

If you want to go hunting this week, the three places I keep

finding broken assumptions are naming and language, architecture

shape, and test style. Our hexagonal setup is deliberately not

textbook DDD, and the tools keep pushing it back toward correct.

Correct for somebody else.

One thing to try:

Name one default your AI coding tool assumes about your domain

that does not hold. Not a hypothetical one. A real one, sitting in

your codebase right now. Then change it deliberately and write

down why.

If you find a good one, hit reply and tell me. I collect these.

Markus