All articles
Engineering
How we think about AI generation
Good AI features start by asking what the user actually wants. Here's how we approach generation in the builder.
Nicolas Daoud
Every team shipping AI features faces the same question: what do we generate, and when? The lazy answer is "generate everything, all the time." The right answer is almost always narrower.
Latency is the feature
A 200ms response that gets 80% of the way feels magical. A 4-second response that gets 100% feels broken. We spend most of our AI budget on cutting latency.
Guardrails beat prompts
Structured outputs and schema validation do more for quality than any amount of prompt tuning. Let the model generate, then enforce the shape.
