Short-form. Between the essays.

Every AI product I've seen ship has crossed the same gap: the demo works in week one; the product doesn't work for six more months.

The gap is not "a few edge cases." It's three orders of magnitude of inputs the demo never saw. The fix isn't more model power. It's a discipline:

1. Log every output that surprised you, ever.
2. Turn each one into a test.
3. Don't ship a model change until the test suite passes.

That's the whole thing. There's no shortcut.

Most of the best product decisions I've made on AI projects have been less interesting than the alternative.

• Use the cheaper model, with a better prompt.
• Add a confirmation step, not an autonomous one.
• Cache it. Ship it. Stop tuning it.

The interesting version is the one founders post about. The boring version is the one customers renew.

I'm trying to be more deliberate about catching myself reaching for the interesting one.

Before I write a single spec, two questions get the team unstuck more reliably than anything else:

1. What happens to the user tomorrow if this doesn't ship? If the honest answer is "nothing," the scope is wrong. We're building a feature, not solving a problem.
2. What's the smallest version that proves it's useful? Not the smallest shippable — the smallest evidence-generating. Often the answer surprises everyone in the room.

Two questions, ten minutes, weeks of saved work.