I have a variant on this. I start with a bullet-list of functionality I think I’ll need. Then I write one test each for items that could impact my architectural choices. Once I finish that first pass I go back and flesh out each item with multiple tests as necessary.

Flushing out the architecture early is the trickiest part for me. Occasionally I miss a seemingly-simple feature that ends up requiring an architecture change. If this happens late in the process, I may pay for the oversight with hours of refactoring.