A student can finish a lesson, understand it in the moment, and still struggle to use the idea a week later. That is the gap between seeing material and actually learning it. It is also why adding more videos, explanations, or questions does not automatically create better SAT preparation.
While redesigning PrepSAT Learning Mode, we found a useful way of thinking about that problem in Matt Pocock's open-source Teach skill. Teach is a set of instructions for an AI agent that turns a folder into a long-running teaching workspace. It is not an SAT curriculum, an app framework, or a component that we installed in PrepSAT.
The Big Idea: Learning Should Have Memory
The Teach approach treats learning as a stateful process. It records the learner's purpose, what they have genuinely demonstrated, which misconceptions were corrected, and what should happen next. Its learning-record format makes an especially important distinction: material being covered is not evidence that it was learned.
PrepSAT's earlier Learning Mode already offered skill primers, worked examples, guided practice, and checkpoints. The new approach keeps those strengths but connects them into a loop that can remember demonstrated knowledge, hint use, confidence, misconceptions, and when a concept should be retrieved again.
What We Adopted
1. Short lessons with one tangible win
Teach asks each lesson to stay tightly scoped and produce a concrete result. PrepSAT now follows a similar rhythm: a retrieval prompt, one concept, a worked example, guided practice with fading hints, an independent application, and a short exit ticket. A student can finish the loop without committing to a full practice test.
2. Evidence instead of exposure
Opening a page does not prove mastery. Neither does earning XP. PrepSAT records evidence only when the student interacts meaningfully with the lesson or checkpoint. Correct independent work, repeated retrieval, hint dependence, and corrected misconceptions all carry different information. Recommendations use that evidence conservatively instead of declaring that one successful answer means a skill is mastered.
3. Retrieval and scheduled review
Teach emphasizes retrieval practice, spacing, and mixing related skills. In PrepSAT, that means sometimes asking a student to recall an earlier concept before showing the explanation again. The learner model can also surface a skill for review because its evidence is old or uncertain—not only because its lifetime accuracy is low.
4. Trusted sources before polished explanations
The Teach skill explicitly warns against treating an AI model's memory as an unquestionable source. We agree. PrepSAT lessons use an approved source registry, beginning with authoritative SAT scope and skill definitions from the College Board's SAT content domains. Sources have review status and provenance, and unapproved entries do not become student-facing lesson content.
5. Reusable learning components
Teach encourages reusable styles, quizzes, simulators, and diagram helpers instead of producing disconnected one-off pages. PrepSAT uses the same principle with controlled React components. Every existing Math and Reading & Writing skill resolves through the same validated lesson structure, while individual skills can still have bespoke explanations, examples, misconceptions, and practice.
How We Changed the Approach for PrepSAT
A personal folder is a useful teaching workspace, but a production platform for students needs different boundaries. We translated the ideas rather than copying the implementation.
A Linear Equations Example
Imagine that Maya can solve one-step equations but often changes a sign without applying the same operation to both sides. A static lesson shows the same primer again. A stateful lesson can do something more useful: ask Maya to retrieve the balance principle, give immediate feedback on the misconception, provide a fading hint on a guided problem, and then require an independent equation before scheduling a later review.
Another student on the same Linear Equations page may need help translating a word problem instead. The route and skill stay the same; the most useful learning action changes because the evidence is different.
The Voice Tutor Learns the Same Boundaries
PrepSAT's voice tutor can use a concise, privacy-safe summary of the current learning context: the active skill, the learner's next review, and relevant misconceptions. It does not receive an unrestricted dump of private history. During an active checkpoint or timed test, it still refuses to reveal the answer and redirects the student toward general strategy or post-submission review.
Credit, Sources, and Further Reading
Matt Pocock publishes Teach within the MIT-licensed mattpocock/skills repository. The clearest primary references are the Teach instructions, its mission format, and its learning-record format. We are grateful for the ideas and equally deliberate about adapting them to PrepSAT's curriculum, security, privacy, and student-safety requirements.