Starting a project
From the problem you state to a plan: the questions Kuyun asks, the kind of work, the brief, and what is yours to do. How the work then moves is How a project's work moves; the overview is Projects.
From your words to a plan
Kuyun asks before it plans. When you start a project it reads what you wrote and asks you a few things it cannot work out for itself — the facts the work is built on that you did not say (where, when, for how long, for whom, what you can eat, what you already know), what you want to end up with, and how much of the work you want it to do for you. It does not guess those and plan on the guess. The questions appear in the project's chat. Answer them and it writes the plan straight away. Leave them and it plans anyway within a couple of hours, using its own best guess for each one; every guess is written down and marked "correct me", so you can change it later and the plan will follow. If you would rather not wait, tell the chat to get on with it.
You say what kind of project it is. You can do that in two places. In the Home chat, ask your assistant — "make me a beginner Python course with exercises" — and it can start the project right there, as that kind of work, asking first for the one or two things only you know when your words leave them open. It describes the project only from what you say there, never from what it remembers about you, since a project can be published. Or it may point you to Create, the better door to look at every kind, attach files, or see Kuyun's questions first. Either way the project is planned and built in the background, and whatever Kuyun had to guess is written down in the project's chat marked "correct me". On Create, starting a project is three short steps. First the kind: every kind is laid out at once as a card with what it is for — a book (bedtime stories, an audiobook, a novel, a memoir), a course (Japanese for a trip, Python from zero, sourdough baking), an article or report (an article to publish, a term paper, a research report, a white paper), a role-play practice (ordering a coffee in Japanese, a job interview, checking in at a hotel), a coding practice (Python interview problems, recursion drills), a library of material people can ask about (the articles on one subject, a reference you consult, a company's manuals, every restaurant in town), a briefing (a morning news briefing, a daily workout plan, a price watch), and last, Custom, for anything else, which tells you how to write it: what you want to come out of it, who it is for, what you already know, and any limits. Two more cards, Music (songs and background music for your work) and Video (short videos for your work), are marked Coming soon: the Kuyun team is working on them, and they cannot be chosen yet. Second, your problem: the kind's card tells you in a few sentences what it makes and how it runs, and you describe the work in a roomy box. Under it, 📎 Add files takes material the work should build on — a syllabus, your notes, a draft: text, PDF or Word (.docx) files, up to 10 MB each and five at most. Kuyun reads the text in them, so pictures, and a scanned PDF with no text in it, are turned away with the reason. (Already have the files — a project you downloaded, or lessons you wrote? import a project (.zip) under the kinds makes it in one step; Projects, "Your project as files".) Third, Kuyun checks: it reads what you wrote before anything is created and asks the few things only you know, each with the answer it would otherwise go with — answer, leave a box blank to let that stand, or let all of its guesses stand — at most twice, and it does not ask for what your files will give; then the project is created, with a name Kuyun gives it from your words. Your files are uploaded first — you see each one go — and a file that does not get through in two minutes, or that Kuyun cannot read, is named, and the project begins without it: add it later from the files beside the chat. The project then opens at once, and its brief and first plan are written behind it, from your files — a minute or two. The page says so while they are, and you can leave it: they are there when you come back, and if the writing fails Kuyun says so and tries again on its own. Each kind is planned and checked the way that work is really done: a book keeps a continuity sheet its chapters are written against; a library keeps every document and every row with its source and date; Custom works a question to an answer with you, finding evidence and weighing it. The kind also decides how the project runs: a briefing looks the same things up afresh every morning and hands you the day's digest — one dated page that leads with what changed and gives each item its source — and a course, a library, a book or a Custom project is built once and kept — a course's readers study it at their own pace. This never limits what you can ask for: anything that fits none of these is Custom. The kind is fixed once the project exists — Kuyun never changes it, and neither does the chat. If you picked the wrong one, start a new project of the right kind; the old one can be archived. The kind also decides how a viewer is met: a course's viewer gets a study partner who teaches it a step at a time, a book's viewer a reading companion who never gives away a story's chapter they have not reached (a work of argument or reference — a history, a science book, a guide — has nothing to spoil, and the companion answers from any chapter and names it), a library's viewer a guide who answers from what the library holds — a document and its passage, a table and its row — and says so.
A book or a library? A book is one work, read in order: it has a first page and a last, and its companion keeps you company along it. A library is many pieces, asked across: essays, stories, poems, recipes, articles, tables — things you look things up in rather than read through. Darwin's Origin of Species is a book; the eighty-five Federalist essays are a library. The same text can be either, and the question to ask is how people will use it: read it, or ask it.
A library is a collection with a purpose — my favourite books, the papers on LLM training, everything by one author — and it holds what you put in it. Material gets into a library four ways, and one library can mix them. Shelve a work: tell the library's chat to put a whole work on its shelf — a Kuyun book, an article, a course, another library; your own, or any published one — with a line saying why it belongs, and, if you like, a section to show it under ("Lessons", "Role-play practice", "Reading"): each section is a folder under Works, in the order you first used it, and works you gave no section sit in Works itself. Nothing is copied: the work stays its maker's, and the library lists it under Works as one item, for as long as it is published, with a mark for its kind: 🎓 a course, 📖 a book, ⌨️ coding practice, 🗣 a role-play practice, 📝 an article or report, 📰 a briefing, 📚 a library; a reader who opens it goes to the work itself, where its own partner takes over. Ask the guide to start a course, a practice or an interview from the shelf and it opens that work for you, at the lesson or scenario you are at, and the work's own partner starts it. In a work opened from a library, ← Back returns to the library. The guide still reads inside every shelved work to answer you. Drop it in: the text, PDF and Word files you upload to it are the collection, read as they are, nothing rewritten — share a file with viewers and the guide can read it. Have Kuyun collect it: say what you want gathered — the articles on a subject, a set of pages — and Kuyun searches, reads and files each one as a document with where it came from and the date it was read. Have Kuyun tabulate it: records of one shape — restaurants, terms, parts, dates — become a table, one row per thing, every value with its source and date. The guide searches all of it at once — the documents, the tables and every chapter of every work on the shelf — reads the passage or the row the search points to, and answers from that: comparing across works and documents, tallying rows, quoting a passage when you ask, and saying which work and chapter, which document or which row an answer came from. When the library does not hold the answer, it says so and names what it does hold nearby. Other projects you link can ask your library the same way.
Works that go together. When you make a course and, beside it, a set of role-plays and a book of stories for the same lessons, tell the course's chat to link them — "link my role-plays as practice, lesson 3 with scenario 3", "link my story book as reading, lesson 3 with stories 11 to 15", "link the next course as the next level". From then on, a viewer who finishes lesson 3 is offered that scene and that story, and at the end the next level. A book can link its sequel the same way. A library can link works on its shelf to each other for the people who open them through it. A linked work is offered only while its viewers may see it, so an unpublished one never shows.
Meeting a viewer at their level. A course, a role-play set, an article or a custom work can carry a placement guide: what to ask someone new, and where each answer starts them — "ask them to read this sentence; if they can, start at lesson 7, otherwise at lesson 1". Tell the work's chat to write it ("add a placement test: three questions, from easy to hard"), or put it in your project's files as PLACEMENT.md beside README.md. From then on a new viewer's first buttons begin with Find my level. Their study partner asks the guide's questions one at a time and starts them where it says. The guide is not among the files viewers can open: only you and the people who may edit the work see it there. Their study partner reads it to ask the questions, though, so a viewer who insists can learn what it says. A placement is a guide, not an exam, so put nothing in it you would mind a viewer reading. Each viewer's level is their own and goes when they leave the work. A library can carry one too, and its answers can start someone in any work on its shelf — "a beginner starts the first course at lesson 1; someone who reads kana starts at lesson 4": the library's guide names the work, and the button under its answer opens that work at that lesson. Beside the button to find where to start, a library with a placement offers From the start, for someone who knows nothing yet and wants no questions: the first work on its shelf, in the order you shelved them, at its first lesson. A library with no placement offers the same first lesson as Start at the head of its buttons, so anyone arriving can begin at once; a library with nothing on its shelf offers only its questions. When the guide lists or suggests works on the shelf, a button to open each one follows its answer.
"Library" is your word for this kind of project, not your viewers'. They meet it by its title: its guide speaks of what it holds, never of "the library", and its buttons say "What's in here?". Write its README the same way, since viewers read it too.
It asks only what only you can answer, and only when a wrong guess would cost something. How the work is laid out — the order of sections, how they are grouped, the format, the naming, the tone — Kuyun decides for itself. Each such choice is written into the goal and context file, marked "correct me", instead of being put to you as a question. Tell the chat to change it and the next plan follows; the order of what is already written, and where in Output it lives, the chat changes at once.
Then it writes the brief: what you want, what will come out of it, what Kuyun will do, and what only you can do. A step is yours only when nobody but you can do it — something in the world, or a decision that is yours and costly to get wrong. Kuyun never hands you its own work to confirm, or a list to supply: it decides, marks the choice "correct me", and goes on. Everything afterwards is judged against that brief, so it is worth a glance — it is the brief file in the explorer, and you can change it like any other document. If changing it raises a question only you can answer, Kuyun asks — and your answer goes into the brief, not just into the next plan. The plan is revised with your answers behind the page — a minute or two; the page says so while it is, and you can leave it. Every question is checked first by a second assistant whose only job is to say whether it really needs you; one that Kuyun could decide itself is decided, and the choice marked "correct me". That check also reaches questions asked earlier, under older rules: within the hour, one Kuyun could decide itself is decided the same way. Leave it and the same thing happens as at the start: after a couple of hours Kuyun answers it with its own best guess, marks the guess "correct me", and carries on. A question Kuyun decides for itself is never dropped: if it came without a suggested answer, it is recorded as "left to Kuyun's judgement", and you can still give one. Nothing Kuyun could decide waits on you. What does wait is what only you can settle: a question only you can answer that came without a suggested answer, and a question Kuyun could not get checked at all. A step that is stuck never waits on you: Kuyun reworks the plan around it (see "When you are stuck or it is").
The brief carries what you said, and nothing else. How Kuyun goes about this kind of work — a course written as a sample and then the rest, a briefing that leads with what changed — is its method, not a requirement of yours, and the brief does not put it in your mouth; a choice Kuyun made for you is listed under "Assumed", marked "correct me". When you ask for something Kuyun cannot make — a recording or an audio file, a file to download, a program run or installed on your device — the brief says so plainly in its "Kuyun does" section and names what you get instead, rather than carrying the wish as a promise every step is then held to.
The brief ends with what the work is about, under "About this work": four short lines saying which language it teaches, or none; the language it is explained in; whether it teaches Python; and whether its practice is done with your hands or body, away from the screen — baking, a knife, an instrument, a sport. Kuyun fills them in, and they decide which of its guidance the work gets. A language course's study partner has you say sentences back aloud. A baking course's sets the task for your kitchen and asks how it went. A Python course runs its examples in the page. A line that is wrong is corrected like the rest of the brief: change it, or tell the chat. Where the work's own practice shows it — exercises you say aloud in a language, coding tasks — that wins over the line. Projects whose brief was written before 10 November 2026 have no such lines, and a course among them keeps all the guidance a course had then.
Kuyun does not give you homework. Your part of a project is only what software cannot do: going somewhere, talking to someone, sitting an exam, deciding something, and telling Kuyun how it went. Finding material, building the flashcards, drafting the documents, keeping the records — that is its work, not yours.
Nothing on the page is yours to tick. Kuyun's steps are completed by its reviewer, a separate model that reads what each step produced and judges it done or sends it back; a step that keeps failing has the plan reworked around it by Kuyun. The reviewer judges a step by what a step can do: when the plan asked it for something no step can make — a file to download, a run of a program — the reviewer says so once and does not hold the step for it; that gap is the plan's to fix, not the step's. A research step is done when the next step can be built on it: when a source could not be reached, it says so, names what stood in, and the work goes on rather than searching again for the same thing. A step the reviewer sent back is changed where the reviewer pointed and kept as it was everywhere else, so the next review reads one correction rather than a new document. Your own steps — the ones marked "yours to do" — you finish by telling the chat how it went, and Kuyun marks the step done with your words beside it.
Kuyun tells you when it is your turn. When the work reaches one of your steps, Kuyun writes to you in the project's chat without waiting to be asked: what it needs from you, why only you can decide it, what you need to see in order to decide (a short list, right there in the message), and how to answer. You also get a notification. It writes once per step, and never in the middle of your conversation: if you are talking to it, it waits until you have been quiet for a few minutes. The work that depends on your step waits for your answer, so nothing is built on a decision you have not made yet. There is no switch to set and no plan to approve: a new plan is judged by Kuyun's reviewer and adopted on its own, and if you disagree with it, say so in the chat and Kuyun proposes another.