Turn what you know into a real course — with a website to match.
Tell your AI agent a topic. Witherspoon walks it through interviewing you, researching real sources, and writing a full course — readings, flashcards, quizzes, and hands-on projects — then builds it all into a fast little website you can share for free. No coding needed.
That example — Same File, Three Addresses — is 4 units, 12 topics, quizzes, interactive widgets, and a printable certificate. All made this way.
What you get
A proper course, not a wall of text.
Every course comes out as readable files you can check and edit, plus a finished website. The whole thing works offline, needs no login, and remembers each learner's progress right in their browser.
Readings that build on each other
One running example threads the whole course, so every topic picks up where the last one left off instead of starting from scratch.
Quizzes that grade themselves
Instant feedback on every question, plus a breakdown of which learning goals to review. No teacher required.
Flashcards & unit tests
Built-in flashcard decks for the key ideas, and a test at the end of each unit with a passing score.
Hands-on projects
Real exercises with step-by-step checklists, grading rubrics, and test cases — including one designed to catch the tempting shortcut.
Everything traced to a source
Every fact and figure links back to a real source with a word-for-word quote, kept in a ledger you can audit. No made-up numbers.
A printable certificate
Learners who finish get a certificate page they can print. (It's honest about being self-reported — more on that below.)
How it works
You talk. Your agent does the homework.
Witherspoon is a recipe for AI agents — Claude Code and friends. You bring the agent; the recipe makes sure it does the job properly.
Say what you want to teach
"Build a course on the water cycle for my 3rd grade class." That's genuinely enough to start.
Answer six quick questions
Audience, size, tone, licence — all in one go, each with a sensible default if you'd rather shrug.
Approve the outline
The one moment it stops and waits for you. Say yes, and it researches and writes the whole thing on its own.
Get a course and a website
Readable files you can edit, a built site you can preview, and free publishing to a public link when you're happy.
The writing phase takes twenty minutes or so — a good moment to make tea. Before it publishes anything, it asks. Nothing goes public without your say-so.
Get started
Two ways in — both free.
Either way, you need an AI agent that can read and write files and run commands on your machine (Claude Code is the obvious one, but anything similar works). That's the whole prerequisite.
Connect the MCP server
Add this URL to your agent as a remote MCP server. It hands your agent the instructions one stage at a time — nothing lands on your machine except your course.
https://witherspoon.up.railway.app/mcp
Then just ask:
Install the skills
If your agent loads skills from disk, one command adds them:
npx skills add colinmollenhour/witherspoon
Or clone the repo and point your agent at it:
git clone https://github.com/colinmollenhour/witherspoon.git
One small install, whenever you get a moment
The website step needs Bun or Node on your machine — your agent will hand you the one-line install command at the right moment, and you can run it while the course writes itself. Skip it and you still get the full course as readable files; the website just waits until you're ready.
Tips
How to get a genuinely good course out of it.
The tool does a lot, but a few habits make the difference between "pretty good" and "I'd actually teach with this."
Gather your materials first
Got a syllabus, slides, PDFs, favourite articles, or your own notes? Put them in a folder and point your agent at it before you start. The course gets grounded in your sources instead of whatever the web coughs up first.
Keep the scope focused
One course should teach one transformation. "Docker for backend developers" makes a great course; "all of DevOps" makes a bloated one. If you're tempted to say "everything about…", split it into two courses.
Take the interview seriously
The six questions at the start steer everything downstream. The one that matters most is audience — "curious 12-year-olds" and "junior engineers" get very different courses from the same topic.
Scrutinise the outline — it's your one checkpoint
Once you approve it, the agent writes the whole course without stopping. Cutting a weak topic at the outline costs ten seconds; cutting it after generation means asking for a rework. Be picky here.
Spot-check the sources ledger
Every figure in the course traces to a row in SOURCES.md with a verbatim
quote. Skim it. If a source looks weak for a claim you care about, say so — the agent will
re-ground it.
Refine in rounds, not one giant pass
Read a topic or two, give feedback, repeat. And if a finished course feels dense or lecture-y, tell your agent exactly that — "this is hard to follow" triggers a dedicated review that reads it like a first-hour learner and proposes cuts.
Preview live while you edit
bun run dev gives you a live-reloading preview — reword a paragraph, save,
and see it on the page. Much nicer than rebuilding to check a comma.
Do the course yourself before sharing it
Take the first quiz. Attempt the first project. Ten minutes as your own student catches the things no automated check can — a confusing step, a project that assumes something the readings never said.
The honest bit
What this is — and what it isn't.
It is…
- Actually free. Open source, no account, no trial, no "pro tier". The course sites are static files, which makes them essentially free to host too — the default publish takes seconds and costs nothing.
- Careful about facts. The agent isn't allowed to use a number it can't trace to a real source, and the build fails if it tries. That's a design rule, not a promise of perfection — which is why the ledger is there for you to check.
- Yours. You pick the licence for your course during the interview, and a course site you publish carries no strings from this tooling — including for commercial use.
It isn't…
- An app. There's nothing to sign up for and no editor to learn. It's instructions your AI agent follows — you supply the agent, and the agent needs file and shell access to do the work.
- A learning platform. No accounts, no dashboards, no tracking. Each learner's progress lives in their own browser, and the certificate says on its face that it's a self-reported record — because that's the truth.
- Instant. The interview takes two minutes, but the research and writing take twenty or more. It's doing real work; let it.
The best way to see it is to make one.
Pick something small you know well. Twenty-five minutes from now you'll have a course website for it.