The receipts

The research behind Learn Breezy

We did the reading so you can trust the method. Here is the actual evidence behind the system, in plain English, and what we built because of each finding.

AI tutoring works when it is designed right

A randomized controlled trial ran a carefully designed AI physics tutor against active in-class instruction. The students using the guardrailed tutor learned more, and they did it in less time. The gains were large by the standards of education research, in the range of 0.73 to 1.3 standard deviations.

What we built because of it: The whole system is guardrailed on purpose. Claude quizzes you and coaches your thinking. It does not hand you the answer.

Kestin et al. 2025, Scientific Reports (Harvard).

Most students use AI in the way that backfires

In a large classroom study, students who leaned on a raw ChatGPT answer engine looked like they were learning while they had it. Then it was taken away for the exam, and they scored about 17% worse than students who never used it. The tool did the thinking, so the learning never transferred.

What we built because of it: Learn Breezy is built to be the opposite of an answer engine. It makes you do the recall, because that is the part that sticks.

Bastani et al. 2025, PNAS.

Two study techniques beat all the rest

Researchers reviewed ten widely used study techniques and rated each on how well the evidence supports it. Only two earned a high-utility rating: practicing recall (retrieval practice, with an effect size around 0.50) and spreading practice out over time (spaced practice). Highlighting and re-reading, the things most students actually do, rated low.

What we built because of it: Those two techniques are the core of the daily loop: recall-first quizzing, then retests spaced across days.

Dunlosky et al. 2013, Psychological Science in the Public Interest.

Testing yourself is the tool, not re-reading

The Huberman Lab episode on studying pulls the practical protocols together. Testing yourself beats re-reading. A gap between an early test and a delayed retest strengthens memory more than back-to-back review. Sleep is when consolidation happens, and interleaving topics builds flexible recall. These are the mechanics the system automates for you.

What we built because of it: The gap-effect schedule, the sleep-aware planning, and interleaved sessions all come straight from these protocols.

Huberman Lab, "Optimal Protocols for Studying & Learning."

It holds up outside the lab

Lab results are one thing, real classrooms another. A World Bank randomized trial in Nigeria found an AI tutoring program raised English outcomes by about 0.23 standard deviations. A separate math tutoring program (Rori) in Ghana reached roughly 0.37 SD, statistically significant. Real, meaningful, and honest about the size.

What we built because of it: It tells us guardrailed AI tutoring transfers to real students, not just study volunteers.

World Bank Nigeria RCT 2025; Rori/Ghana field trial.

The protocols we automate

The practical mechanics come from the Huberman Lab episode on studying, which pulls the learning-science protocols into one place. The system turns each one into something that just happens on schedule.

Testing beats re-reading

Pulling an answer from memory does more for learning than reading it again. That is why the default study action is a short-answer quiz, never a summary.

The gap effect

An early test followed by a delayed retest, a day or a few later, strengthens memory more than reviewing back to back. The system schedules that gap for you.

Sleep consolidates

Memory settles during sleep, so sessions are spread across nights on purpose. No all-nighter cram that skips the step that actually locks things in.

Interleaving

Mixing topics and question types in one session builds flexible recall, the kind you need when the exam does not tell you which method to use.

Source: Huberman Lab, “Optimal Protocols for Studying & Learning.”

The honest part

Here is the honest version. Independent, at-scale ed-tech effects usually land somewhere around 0.05 to 0.20 standard deviations. Nobody can promise you a grade jump, and we will not. What we promise is different: we teach you to use AI dramatically better, and we build your studying on the two techniques the research actually backs. The work is yours. We just make sure the work is the right kind.

Connect your Claude