Before, build, after: how I connect lessons

I am developing a system that helps me prepare materials and keep a record of what happens in class. It needs to stay flexible as my teaching evolves.

It runs in three phases around every lesson: Before, Build, After. Information from one phase informs the next.

Before

The system assembles what it already knows: the student's profile, past session logs, and trackers — grammar, pronunciation, a spaced-repetition queue, a ranked list of their frequent errors. From that it writes a focused brief: what to recycle, which errors to target, what's unfinished.

Build

Technical. Markdown files hold the memory; scripts manage review schedules, rank priorities, and check lesson structure, links, and consistency. The system produces interactive HTML lessons for live Zoom sessions and mobile-first homework, with neural-voice audio. AI helps draft materials, analyze transcripts, and shape feedback.

Before a new lesson is shared, separate AI reviewers check its language, pedagogy, task logic, and audio alongside the automated checks. I review the preview, authorize the full review, and approve the final materials before publication. These checks help find problems; they support my judgment about what suits the learner.

Pedagogical. I document teaching principles drawn from language-learning research, established practice, and my own lessons. I use them to guide preparation and review. For example:

• Recognizing a word and using it independently are different skills. If the goal is speaking, the practice needs opportunities to produce the word.
• For a familiar task using language already introduced, an initial attempt before the model can show what the learner recalls independently.
• A word the student reached for but couldn't find comes back next lesson — scheduled, not left to memory.

The machine creates the draft; I decide what stays, what goes, and what changes.

After

This is what makes it a system, not a generator:

• A short debrief — what we covered, what was hard, new vocabulary, next focus, homework.
• With the learner’s agreement, transcript analysis of the Zoom recording — abandoned sentences, missed errors, the first unprompted use of new language, plus a look at my own teaching.
• Proposed updates to the profile, vocabulary log, spaced-review schedule, and trackers. I review the evidence and approve changes before they are applied.

That updated state feeds into the next lesson, so each one starts where the last ended — not where my memory left it.

The before-and-after loop matters far more than the generation in the middle.

I won't claim this beats a disciplined teacher with a good notebook. It helps me maintain that discipline, with review points built into the workflow. And the takeaway holds even if you build nothing: getting AI to produce a lesson is easy. Capturing what actually happened and letting it shape the next one is where the value is.