Feb 2026
Stride
VisitAI-powered focus companion
Stride is a focus companion that learns how you work and steps in before you get distracted. It tracks signals during a session — task type, time of day, session length, how you've responded to past nudges — and uses Claude to predict the moment you're about to drift. Then it intervenes a few minutes earlier than that.
The core idea is pattern recognition. The more you use Stride, the better it gets at catching you. If you consistently lose focus 25 minutes into afternoon sessions, it learns that and intervenes at minute 22 next time.
I built it solo, end-to-end. Next.js, Supabase, Claude via the Anthropic API, and Opik for intervention evals. Below are the decisions that took the most thinking.
Key decisions
Predict, don't react
Most focus tools react after you've already drifted. Stride does the opposite — it watches signals during a session (time of day, task type, session duration, how you responded to past nudges) and tries to predict the moment you're about to lose focus, then intervenes a few minutes earlier. Claude is the reasoning engine. The more sessions a user runs, the sharper the predictions get.
Pattern recognition as the core feature
If a user consistently loses focus around 25 minutes into afternoon deep work, Stride picks that up. Next time they're in a similar session, the intervention lands at minute 22 instead of minute 30. The product gets more useful the longer you use it, which is the only kind of retention loop worth building for a focus app.
Evals before shipping interventions
Every intervention runs through Opik for evaluation before it reaches a user. LLM-as-judge scores each nudge on helpfulness, timing, and tone, and the user's response (engaged, dismissed, ignored) feeds back in. I can ship prompt changes and watch the eval scores move within a day, not a week.
Activation before retention
Early data showed users starting a session and never receiving an intervention before they dropped off. The fix wasn't a bigger feature — it was earlier evidence. I shipped a first-session overlay and a hardcoded demo nudge so every new user gets the AI moment in their first session, not their fifth.