CourseStudio

People who know a subject well are usually bad at turning that knowledge into a course, because the work is long and boring rather than hard.

A teacher, a developer, or a domain expert already has the raw material sitting around: lecture notes, internal docs, a workshop transcript, saved articles. Turning that pile into something a student can learn from means reading all of it again, deciding which ideas matter, working out the order to teach them in, writing every lesson, drawing diagrams, writing quiz questions, writing the answer explanations, and designing a final project with a marking scheme. For a twelve lesson course that is weeks of work. Most people never start.

The AI tools that claim to fix this do not fix it. Ask any chatbot to "make a course about X" and you get a list of chapter titles. That is a plan, not a course. Nobody can learn from a plan. The gap between a table of contents and a course a student can actually take is where all the real work lives, and that is exactly the part nothing automates.

This hits three groups. Independent teachers and creators lose weeks of unpaid work before they can publish anything. Engineering and support teams write onboarding docs that new hires skim once and forget, because a document has no questions, no feedback, and no way to check whether you understood it. Subject experts who would be excellent teachers never become teachers, because the production work sits between them and their first student.

We are not solving "generate text about a topic." We are solving the step after that: turning knowledge someone already has into a finished learning product.

CourseStudio takes raw knowledge and produces a complete interactive course. Not an outline. A course a student can open and start taking in the same minute it was generated.

You drop in your material as pasted text, a transcript, a .txt, .md or .pdf file, or a URL. You type one plain sentence such as "turn this into a beginner friendly course that teaches React from scratch." You press Build Course.

The pipeline runs in front of you instead of behind a spinner. On our bundled demo it reads 18,863 words across six documents, pulls out 42 concepts, maps 72 relationships between them including 47 prerequisite links, merges passages that repeat the same idea, then designs 5 modules and 12 lessons. The knowledge map draws itself on screen while this happens, so you watch the system work out which idea has to be taught before which other idea. That ordering is the part a chapter list never gives you.

Then it writes the actual course. Every lesson is built from typed content blocks rather than a wall of markdown: explanations, key concept cards, worked examples, code blocks with copy buttons, comparison tables, step sequences, callouts, and a generated diagram. There are five diagram types, drawn as real SVG components from structured data, so a flow diagram, a process diagram and a concept map each render correctly instead of being a picture of text.

Each lesson ends with a quiz that grades itself. Sixty questions across the demo course, mixing multiple choice, true or false, short answer and scenario questions. Submit and you get your score instantly, with the correct answer highlighted and a written explanation for every question. Each lesson also carries a hands on exercise with steps, hints and success criteria. The course ends in a final project, "Build Your First Tool-Using AI Agent", with 8 requirements, 8 steps, deliverables, four stretch goals, and a five criterion marking rubric weighted to 100 points.

The creator can edit anything. Click any block and edit it. Drag lessons into a different order. Regenerate a whole module. Use Improve with AI on a lesson or a single block with seven actions: make it simpler, make it more advanced, add an example, add an analogy, add an exercise, shorten, expand. Edit a question in the assessment studio and the learner sees the change immediately.

Then press Preview as learner and the whole thing flips into student mode. Sidebar navigation, reading progress, working quizzes, progress that saves. Press Publish and you get a real route, /course/ai-agents-masterclass, that loads on its own and survives a refresh.

Two things we care about most.

First, it never needs an API key. There are two AI providers behind one interface. RealAIProvider calls Claude through a server route using Zod structured outputs, so the model's output is schema valid or it gets rejected. DemoAIProvider runs entirely in the browser with no network at all: it segments sentences, extracts n-gram concepts weighted by how often a term appears, whether it sits in a heading, whether a sentence defines it, and how many documents mention it, then infers prerequisites from first mention order and paragraph co-occurrence, clusters concepts into modules, composes lessons from real source sentences, and generates quiz distractors from neighbouring concepts. Paste text about photosynthesis and it builds a photosynthesis course offline. The demo cannot fail on stage because of a network or a rate limit.

Second, nothing in the app is a dead button. Every external capability has a working fallback. If a URL will not extract, a paste panel opens in its place. If the model is unavailable, the local engine takes over silently. Diagrams are structured SVG rather than an image API. Storage is localStorage rather than a database. Analytics are clearly labelled as a seeded cohort.

Stack: Next.js 16 App Router, React 19, TypeScript in strict mode, Tailwind v4, Zustand with persistence, React Flow for the knowledge map. Nineteen routes, clean production build.

We tested it the way we would test a product we had to ship. Nine Playwright tests. One of them walks the exact stage demo click by click and fails if a single error appears in the browser console. We also ran that whole flow against the deployed Vercel site, not just localhost: build, knowledge map, diagrams, publish, quiz scoring and progress after a hard reload all pass live with zero console errors.

On the quiz being honest: our test script picks the first option on every question and scores 0 of 5. It is really grading, not pretending.

Built with Claude Fable 5.1, which did the architecture, the full heuristic engine, the twelve authored lessons, the UI, and the test suite.

RealAIProvider calls Claude through a server route using Zod structured outputs, so the model's output is schema valid or it gets rejected. Built with Claude Fable 5.1, which did the architecture, the full heuristic engine, the twelve authored lessons, the UI, and the test suite.