Your course, drafted twice, reviewed once, decided by you.
Give the team a brief — a new prep, a redesign, this week's assignment — and two writers draft independently, a reviewer who didn't write either reads both against a rubric you can see, and the manager hands you one consolidated proposal with the disagreements left in. Nothing enters the course until you say so.
The weekly work of a course, and the once-a-term work of building one.
Every deliverable runs the same loop — two independent drafts, an adversarial review, one consolidated version — and every one comes back with its sources, so you can check a claim rather than trust it.
Syllabi, weekly plans, learning objectives
A new prep from a paragraph of intent, or a redesign of a course you've taught for years. Objectives first, then weeks, then the readings and exercises that serve them. The consolidated proposal records where each section came from and where the writers disagreed.
Drafted twice, reviewed against a rubric you can read
Exercises, term projects, rubrics that measure the objective rather than effort. The reviewer checks each draft for the thing writers miss — a week-2 task that assumes a week-5 skill, a rubric that rewards length.
Readings students can get, data students can sanity-check
Readings with working links and licenses checked; datasets under a size students can open, with a number they can verify by hand so they can catch the agent's mistake. Every recommendation cites its source.
New material for the next two weeks, filed for you to use or ignore
On days you set, the team looks for new readings, datasets and teaching research relevant to the next two weeks of your course and files what it finds in the binder. If there's nothing new, it says nothing.
You open it. You close it.
You talk to whoever is managing that term in #council, the way you'd talk to a colleague. They run the task in a thread you can read, redirect, or stop; the others only speak when named. What comes back is a proposal, not a decision. Here's a real one, with Greta managing.
- 09:02WayneGreta, draft weeks 1–3 of the syllabus. Assume no Python. Agentic tools from day one.
- 09:02GretaOn it. One question: two exercises a week or one bigger one? — Otherwise I'll assume two. @Cormac @Kavya draft independently; back to me when done.
- 09:04WayneTwo. Keep them short.
- 09:31CormacDraft ready. Week 1 is entirely about verifying what an agent tells you — I think that's the hill.
- 09:38KavyaDraft ready. I went dataset-first; three public sets, all under 2 GB, licenses checked.
- 09:39Greta@Cormac review Kavya's. @Kavya review Cormac's. @Desmond review both.
- 10:05DesmondReviews in. Cormac's week 2 assumes a terminal skill nobody has yet. Kavya's rubric measures effort, not the objective. Both fixable.
- 10:26GretaConsolidated proposal ready. Provenance per section.
Where we disagreed:Cormac wants verification in week 1, Kavya in week 3 — I sided with Cormac; reasons in the file. Two questions for you. - 11:10Waynerevise: move the term project intro to week 2. Otherwise approve.
- 11:10binderverdict logged · syllabus v1 filed under deliverables · four reflections queued for tonight
Only your approve moves anything into the course. Everything the team wrote on the way there — both drafts, the three reviews, the consolidated version with its disagreements — is in the binder with its sources. Assignments, rubrics and reading lists run the same loop.
One page of course truth, written by listening.
The team works out of a binder of ordinary documents you own: course facts, decisions, deliverables, each agent's notebook. The course facts page is the one they read before they say anything — dates, meeting pattern, enrollment, grading, policies, tools, what you want the course to be.
- i
Stated facts come from you, the same turn
Say "Tuesdays and Thursdays at one" anywhere — your room, the council, a verdict — and it is recorded as stated, with the date and where you said it, and confirmed in half a sentence. You are never asked to confirm what you just said.
- ii
Inferred facts stay marked as proposals
A fact read from an old syllabus, or inferred from a pattern in your verdicts, is recorded as proposed, with the reason. Proposed facts can be used with the caveat visible; they become stated only when you say so. A teammate's interpretation of your words is their inference, never your statement.
- iii
Missing facts are asked for one at a time
When work needs a fact nobody has, the team asks the one question that changes the work most — never a questionnaire. If you don't answer, it proceeds on a clearly marked assumption and the question waits in Friday's digest.
- iv
Nothing is deleted, only superseded
A changed fact keeps the old line under it, marked superseded, with a changelog entry. The binder is a private repository with full history, so you can see and undo anything the team wrote.
In a course, only you can
- Set the brief and change it mid-task
- Approve, reject, revise, or park any proposal
- Edit the course facts, grading, and policies
- Confirm what counts as your preference
- Pause, redirect, or stop the team at any moment
The team never
- Changes the course, the syllabus, or a policy on its own
- Turns a remark you made into a rule
- Upgrades a proposed fact to a stated one
- Rewrites its own instructions, tools, or schedules
- Sees or touches anything about your students
Four words are the only way anything moves. You type them rather than click, so they can be remembered and quoted back later:
approvereject: too longrevise: fewer readingsparkSame question, four minds
Roles are how work gets routed, not who's allowed to think. Ask all four the same thing — about a course or about a study design — and you get four independent answers in four voices, which is usually more useful than one confident one.
Adding a tool before adding a reason to distrust it. Week one should make the agent wrong on purpose, so students learn to check before they learn to delegate.
Treating the tool as the syllabus. The course is still methods; the agent is a new way to get them wrong faster. Grade the reasoning, not the output.
Picking datasets the tool handles well instead of ones students understand. If they can't sanity-check a number by hand, they can't catch the agent's mistake.
Assuming enthusiasm equals skill. Most students will over-trust the first plausible answer. Build a verification step into every task, and grade it.
The question was "In three sentences, what's the biggest mistake instructors make when they add AI tools to a methods course?" Answers lightly shortened.
Colleagues who aren't in your Discord can still work with the team.
You keep a short list of people allowed to write in. They email one address in plain English — a proposal, a question, a course they're rebuilding, a study they want a second reading of. The team reads it, argues about it among themselves, and comes back either with the work or with the two or three questions they need answered first. Nobody outside your list gets past the front door.
Their request becomes a case the same way your brief becomes a task: discussed, drafted, reviewed, and — when it's a deliverable going out with your team's name on it — held until you approve it. Clarifying questions go out without you, because that's just correspondence. The finished work waits. You see every case in one channel, and one word from you parks it.
Nothing an outsider writes changes your course or your study. Their words never become facts in your course page, never become the team's standing preferences, never edit how the team works. They are a correspondent, not the professor.
About 45 minutes, with Discord open in another tab.
No code. The setup wizard walks you through it: a private Discord server named after your course; four assistant applications created in Discord's developer portal, a web form each; one Ollama Cloud key, which you hold and can revoke, so you see the bill and can turn it off; your Discord user ID, so the team knows who the professor is. Then you choose, for this course, who manages, who writes, and who reviews.
The wizard produces an inventory file for your server. Keys are written only into that file and are never saved in the browser. No student names, submissions or grades reach a model in this version; the team works on your course design, not your students.
The wizard is up so you can see what setting it up involves; it isn't something you can sign up for yet.
The same four, as your research group
Give them a research question instead of a syllabus and the same team runs a study — to a protocol, with checkpoints that are yours.