Architecture documents that actually get read

The cheapest place to change a system is a document. The most expensive is production. Everything in between is a matter of degree.

We keep the structure deliberately short: the problem in one paragraph, the domain model, the boundaries and who owns what data, the failure modes, and the decisions we considered and rejected. That last section does more work than the rest combined — six months later it answers 'why is it like this?' without requiring the original author.

Length is not rigour. A document that nobody finishes reading has failed at the only thing it was for.