Symptoms
The AI assistant doesn't intervene when you visibly drift, or the drift log is full of "unclear" verdicts. You expected it to catch you scrolling Reddit during your study session and it didn't.
This is by design
The classifier is instructed to prefer "unclear" over a guess. Only high-confidence distracted verdicts trigger intervention. False positives — force-quitting your IDE because the screen looked off-task — destroy trust faster than any other bug class. Missing a real drift is recoverable; force-quitting a legitimate workflow is not.
Tune your taskContext
The single biggest lever is the task description. Vague tasks produce vague verdicts.
- Bad: "work". The classifier has nothing to anchor on — almost anything could be work.
- Better: "writing the Q3 marketing email". Specific verb, specific artifact.
- Best: "writing the Q3 marketing email in Gmail or Notion — anything else is off-task". Tells the model what's allowed AND what isn't.
If it's still not catching you
- Verify Screen Recording permission (see the dedicated troubleshooting page) — without it the vision input is empty and the classifier degrades to text-only.
- Install the browser extension. The DOM text excerpt grounds the classifier when the screenshot alone is ambiguous (e.g. a YouTube tutorial that's actually relevant vs. an entertainment video).
- If the block is on Nudge, the first drift gets a 5-second warning and then force-quits if you keep drifting; Hard mode force-quits on sight. If you want every drift killed immediately, switch to Hard.
- If you're consistently catching false-negatives on the same site, that's a tuning issue, not a permission issue — re-write the taskContext to explicitly call it out.
Still stuck? Contact support at hi@focusdragon.app with your taskContext and the drift-log entries you'd expect to be "distracted" but aren't.