BLOCKTARIOPeople ask how a small team ships as much as BlockTario does. The honest answer is that we don't write most of our code by hand anymore — we direct it.
Here's the actual workflow, not the marketing version of it.
We don't pick one tool. We use both, on purpose.
Claude handles planning, larger implementation passes, refactors, and code review. Cursor handles the moment-to-moment work — inspecting diffs, making small UI adjustments, the kind of visual back-and-forth that's faster with your hands on the keyboard. Both operate on the same repository, so neither tool works in a silo.
Every feature starts as a written spec, not a prompt.
Before any code gets written, we write a short spec: the user story, the requirements, and the acceptance criteria. Something like:
As a customer, I want my progress saved automatically so I can leave and come back without losing work. A response should survive a page refresh. A user should never be able to read another organization's data. Relevant tests should pass.
That spec goes to Claude first — not to implement, just to plan. We ask it to inspect the existing codebase, propose an approach, flag security considerations, and list what it's assuming. We review that plan before a single line changes.
Nothing merges without a second, adversarial review.
Once a feature is implemented, we open a separate session and ask Claude to review the branch as a skeptical senior engineer — specifically hunting for broken permissions, data leakage, unvalidated input, and anything that looks like it works locally but wouldn't survive real traffic. That review happens before we look at it ourselves, because a machine that isn't emotionally attached to the code it just wrote catches things we'd talk ourselves out of.
Repository rules keep the AI honest.
We keep a running set of ground rules in the repo itself — a running instructions file the AI reads before touching anything. Things like: never invent data that wasn't provided, validate every AI output against a schema, keep sensitive calculations outside the model, and don't refactor code nobody asked you to touch. It's the difference between an assistant that helps and one that quietly rewrites half your app because it thought the code was "cleaner" a different way.
None of this makes the AI infallible. It still gets things wrong, and we still read every diff. But it means two developers with the right process ship at a pace that used to take a much bigger team — and it's the same process behind everything from Founder OS to what we're building next.