Why an LLM can't replace your Architecture Decision Records

"I don't need this — I can just ask ChatGPT to generate the architecture."

I hear this a lot. And it's true: any LLM can produce a C4 diagram in seconds. But always, there is a catch.

Here's what it can't do:

It doesn't know that your team tried microservices eighteen months ago and burned six months on service mesh complexity before reverting to a modular monolith.

It doesn't know that your regulator flagged a specific data residency concern that ruled out three otherwise obvious vendor choices.

It doesn't know that your lead architect is leaving in Q2 and you need decisions externalised before institutional knowledge walks out the door.

And it doesn't tell you to which context to add, which is irrelevant.

An LLM generates plausible structure with other people's decision inherited and not even spelled out.

An ADR records your decision, with your constraints, for your future team.

The value of a Tech North Star isn't the diagram or the nice slides. It's the reasoning behind it — human-authored, context-specific, and designed to guide the next decision.

Use the LLM to move faster inside the structure. Don't use it to avoid creating a structure.

The structure, written by a human who has had to live with the decisions: technorthstars.hu.