The easiest AI to build is one that always has something to say. The hardest is one that knows when silence is the better answer. When we started building the nudge delivery system, our first prototype was enthusiastic to a fault — it would remind you of everything, all the time, with the persistence of a golden retriever who found a tennis ball.
The Over-Nudge Problem
In early testing, users started disabling notifications within 48 hours. Not because the reminders were wrong — they were technically accurate. The problem was volume and timing. Getting reminded about six tasks in a 30-minute window makes the whole system feel like spam, regardless of how relevant each individual task is.
We learned an important lesson: correctness is necessary but not sufficient. A notification can be factually correct — yes, you do need to review that document — and still be counterproductive if it arrives at the wrong time, in the wrong context, or alongside too many other reminders. The user experience of a notification system is not about individual accuracy. It is about the aggregate pattern.
Teaching Restraint
We introduced per-task daily limits, global cooldowns, and cross-task prioritization. When multiple tasks compete for the same window, the system now picks the one with the highest urgency-to-effort ratio and holds the rest. The goal shifted from maximizing reminders to maximizing completions.
The cooldown system works on multiple levels. After delivering a nudge, the system waits at least 15 minutes before sending another — regardless of how many tasks are pending. After a task is completed from a nudge, it extends the cooldown further, because the user just demonstrated productive behavior and does not need to be interrupted.
The Quiet Score
We developed an internal metric we call the Quiet Score. It measures the ratio of delivered nudges to dismissed nudges over a rolling 7-day window. A high quiet score means the system is mostly delivering nudges that get acted on. A low score means it is being noisy.
The quiet score feeds back into the delivery algorithm. If a user is dismissing a lot of nudges, the system automatically reduces frequency and becomes more selective about which tasks to surface. Over time, it converges on a pattern that feels natural rather than intrusive.
What We Learned
The biggest lesson from building this system: AI restraint is harder to implement than AI capability. Anyone can build a model that generates a response for every input. Building one that evaluates whether a response is warranted — and stays silent when it is not — requires a fundamentally different design philosophy. We are still refining it, and probably always will be.