PrepaTutor AI

I started by watching how each side of a prépa actually uses AI today. Students revise against the clock with generic chatbots, and the failure is structural: a topic that branches, one question opening three more, collapses into a single scrolling thread they cannot find their way back through the next day. The tool answers, but nothing accumulates.

Teachers carry the mirror image of that weight. Every course, exercise sheet and quiz is hand-built at a level that has to hold against the concours, and the only feedback loop is the marks, weeks later. Both sides needed AI, but not the same AI, and not in two separate products.

The suite is framed as three products behind one shell: study boards where revision happens, a document studio where material is authored, and a progress layer that reads how students actually work. The shell is a launcher: a calm grid of tools under a greeting, wearing the same glassy floating header everywhere.

Because every tool is a permission flag rather than a menu entry, the launcher shows each person exactly what they may use and nothing else. Four people use four different versions of the product, and each of them sees a complete product, not a restricted one.

The riskiest design question was the student tutor. The obvious shape was the one every AI product already has: a chat. I explored three directions: the thread everyone knows, a living document that rewrites itself as you ask, and a spatial board where every exchange is a node on a canvas.

Walking a real revision session through each shape settled it. The thread collapses branching: by the third follow-up, the second question’s context is gone off the top of the screen. The living document hides the path: you see the answer, never how you got there. The board was the only shape where branching is free and the way back stays visible, so a session becomes a map a student can stand back from.

The first boards proved the direction and exposed its cost: past a dozen nodes, a map of identical cards is as illegible as the thread it replaced. The second version answered with a grammar. Every node has a type, question, explanation, hint, exercise, document or fiche, and every type has a fixed colour, so a dense board reads by shape before a single word.

The spatial map also gained a linear twin: a reading panel that walks the same thread in order, for the moments when a student wants to be led rather than to explore. And a tabbed workspace groups chats, documents and resources, so the tutor always has the student’s own material one tab away.

The studio’s design question was trust: a teacher will not hand students AI material they cannot stand behind. So generation is parameterised by the real prépa ladder, the filière (MPSI, PCSI, MP, PSI and the rest), the niveau from Sup to Spé étoile, the matière and a difficulty that climbs to concours level, and every request can carry a reference PDF. The prompt speaks the teacher’s language before the model writes a word.

The output opens in a rich editor with an assistant beside it, because generated is a starting state, not a final one. Ask for another example, a harder exercise or a shorter proof and the document rewrites itself. The maths is rendered live and printed straight from the page, so the formula in the editor is the formula in the exported PDF.

The progress layer began from a distrust of scores. A right answer can be a lucky guess; a wrong one can hide honest work. So the model reads behaviour instead: how long a student waited before opening a hint, whether they revealed the solution straight away, how often they came back, how well they said they understood. From that stream it derives mastery, learning quality, autonomy and a struggle index, per student and per concept.

The same data is read at three altitudes. A student sees their own progress and what to work on next. A teacher sees their classes. A supervisor sees the whole school as a risk matrix with an alert feed. And the analytics surfaces deliberately step out of brand orange into a quieter, health-app palette, because dense numbers need calm, not energy.

Three tools that could have felt like three apps are held together by one language. A single warm orange carries every action, Geist sets every word, and a faint graph-paper grid gives the suite a studious ground without a single illustration. Surfaces are glassy cards floating over that grid, and everything is drawn in light and dark from one set of tokens, so neither theme is an afterthought.

The colour is disciplined: orange only ever means action, while node types and the analytics layer keep their own fixed palettes, so a hue always carries meaning. And when someone is lost, an in-app assistant does not just describe where to click, it dims the page and cuts a spotlight around the real control on their own screen.

PrepaTutor AI shipped as three AI tools behind one role-aware launcher, drawn in light and dark from one token set. What I keep from it is the board: giving an AI conversation a spatial structure did more for comprehension than any prompt tuning could, because it changed what a session leaves behind.

The other lesson is about trust surfaces. A teacher trusts the studio because its parameters speak prépa; a student trusts the tutor because its context is visible on the canvas; staff trust the scores because the pipeline behind them can be explained. On an AI product, most of the design work is the work of making the system inspectable.