Efficiency mode exists so Brainback can be a full AI assistant, not just a tutor. Users draft emails, brainstorm, summarize articles, and get help with life admin. All the things a normal AI assistant does.
But Efficiency mode has one job that a normal assistant doesn't: catch academic problems in the user's declared subjects and redirect them to Learn mode.
The classifier is a Haiku call on every message. Input: the message and the user's learn_subjects. Output: a JSON object with is_academic_problem, subject, and reason. If is_academic_problem is true and the subject matches a declared subject, the message is redirected — Efficiency mode never sees it.
The design tradeoff is asymmetric. A false negative — an academic problem slips through to Efficiency mode — is bad but recoverable; the normal-assistant prompt still refuses to solve declared-subject problems. A false positive — a non-academic message gets redirected to Learn mode — is worse; it interrupts the user's task with a nag that doesn't apply.
We tuned the classifier to prefer false negatives. Precision matters more than recall. If Efficiency mode is right most of the time and wrong occasionally, users trust it. If it wrongly redirects one message in ten, they turn it off.
Written by Brainback Research. Published Apr 27, 2026. Filed under Product design.