How Chameleon teaches

Before writing a word of tutorial, Chameleon takes the chapter apart. The text is broken into atomic facts, each tied to the place in your source it came from and tagged with the kind of thinking it asks of you, from plain recall up to application. Those atoms are then fused into a short list of concepts sized for one sitting, deduplicated, and put in the order your book presents them. The tutorial follows your source's own narrative arc rather than an outline the model invented.

That anchoring is also what makes the citations work. Because every concept knows which pages it was built from, every card can point back to them, and anything the model writes that cannot be grounded in your material is checked and dropped before you see it.

The reader shows one concept per card. A spine on the left lists every card in the chapter so you always know where you are and can jump anywhere; the margin on the right carries the gloss and the source citation for the card you are reading. Arrow keys turn the pages. A focus mode strips the frame away, and the reading settings (typeface, spacing, colour overlays) follow you in.

The point of the card is pacing. A long scroll invites skimming; a card asks you to finish one idea before you get the next. Long chapters arrive in parts with distinct jobs, so part two deepens what part one oriented instead of restarting the lecture, and each later part opens with a recap of where you left off.

A tutorial that only defines things produces recognition, and exams demand more. So alongside the concept cards, Chameleon plans dedicated connection cards for each chapter:

The targets are chosen per chapter by a dedicated planning pass, from what your material actually says. Concept cards open concrete, with the case or the number first and the definition after, and later cards call back to their neighbours so the chapter reads as one argument rather than a glossary.

Chameleon identifies the discipline your files belong to and writes in the persona of a specialist in it. That is more than tone: a law chapter is held to the standards a law examiner applies, an economics chapter gets the marginal reasoning an economist would insist on, and the misconception cards draw on the errors that field is known for. The same persona marks your written quiz answers and explains what to correct.

Generated prose is also held to a house style. The tics that make AI text recognisable are banned outright, so what you read sounds like a tutor who knows the material, without filler.

Figures in the tutorial can open with a one-line prediction prompt: which way does the curve shift, what happens to the equivalence point, which region rejects the null. You commit to an answer, then reveal. Guessing first and being corrected is one of the most reliably useful moves in learning research, and here it is built into the act of reading rather than saved for a quiz three days later.

Once you have reviewed cards and sat quizzes, the reader reads your record back. Concepts your history marks as weak get a dot in the chapter spine and a REVISITnote in the margin, and their prediction prompts open ready to test you again. When an explanation does not land, one press of “explain differently” asks the persona to take another run at the same concept from a different angle, on the spot.

Every chapter arrives with flashcards and a quiz built from the same grounded concepts. The flashcards are enrolled in an FSRS-5 spaced-repetition schedule that models the memory of each card and decides when it is due; the quiz mixes multiple choice with written answers and marks both, with specific feedback. What you miss flows back into the review queue and into what the reader flags, so the three surfaces study the same ledger.