Neo's question for the team and Lan. The theory week is done: the student knows what an AI agent is. Now they touch the tools for the first time. Do they start with our one box, or with the terminal, Claude Code and Codex set up by hand? And if students set up their own memory, as in Lan's class, does it clash with the course's record? Try the first lab yourself first: course.neotodak.com/demo.
| Min | The student | What they learn without being told |
|---|---|---|
| 0–5 | Opens the Workbench. The game is already built. Plays it. The right paddle plays by itself; theirs does not move. | A game is a thing you can hold and change, not a thing you admire. |
| 5–15 | Reads the desk's greeting, types "my paddle doesn't move". Sees "Build desk · Claude Code, because you asked how the game behaves." The game rebuilds itself. They play again. | There are two helpers; one changes behaviour, one makes pictures. The box chooses and says why. |
| 15–20 | The desk asks what changed. They write one sentence. | Understanding is part of the job, not an afterthought. |
| 20–35 | Types "draw me a cool title screen". Sees "Design desk · Astra". A picture lands on the design board. Then "make the ball faster". | The same box, two very different kinds of work. They start guessing the route before the box says it. |
| 35–45 | Types "build me the whole game". Gets the step list and "not in one go". | The course is steps, and the box will not let them skip. |
| 45–60 | Opens the Code tab, reads the diff of their paddle change with the lecturer. Opens the Journey tab and sees everything they typed. | Code is readable. Prompts are on the record, and quality will be judged. |
| Box first (recommended) | Hard way first | |
|---|---|---|
| Day one feels like | Playing and asking. Everyone ships a change in the first hour. | Installing, signing in, learning two tools' habits. Some ship nothing on day one. |
| What protects "not vibe coders" | The rules inside the box: one step at a time, explain it back, prompts graded, no-AI drills from week 4. | The friction itself. It filters people, it does not teach them. |
| Who it suits | Beginners who came to make a game. The tool is the medium. | People who came to learn Claude Code and Codex as tools, like Lan's "Zero to Clauder" class. |
| When the raw tools appear | Week 2, one supervised look "under the hood". Required from week 4. | Day one, all of them. |
| Recording | Complete, the box records every request itself. | Depends on hooks and transcript collection on each machine. |
| Risk | A student leans on the box and stalls at week 4. Answer: the drills and the walkthroughs are pass or fail. | Early dropouts, and a lecturer's first weeks spent on installs. |
| Fast lane | Anyone who finishes a step early may open the raw panels. | – |
| The student's memory (project brain) | The course record (journey) | |
|---|---|---|
| What is in it | Their decisions, conventions, bugs, what they tried. | Every prompt, route, build, review, publish, with time and student. |
| Who owns it | The student. They keep it after the course. | The Academy. Assessment evidence, under the consent they signed. |
| Who reads it | The AI, every time it helps them. | Lecturers and the student. The AI never reads it. |
| Where it lives | Today a file in their game. In Lan's exercise, their own pgvector database, one per student, plugged into the same remember and recall commands. | One table in the Academy's database, append-only. |
| Can they clash | No: different content, different readers, different owners. The only design that clashes is one shared memory for all students, where one person's decisions leak into another's AI answers. We never build that. | |