Everybody’s first attempt at how to create an online course with ChatGPT goes the same way, and it works right up until it doesn’t.
You paste in your material. ChatGPT writes good lesson content. You copy it out, open your course tool, make a lesson, paste, reformat, notice the quiz questions are generic, rewrite two of them, and repeat. By lesson four you’re doing data entry with extra steps, and the AI’s contribution is starting to look like the cheap part.
There are two methods here. The first is the one you already know, written down honestly with its ceiling marked. The second removes the copying entirely.
Try Mini Course Generator free – connect ChatGPT and get a published course instead of a transcript.
Key Takeaways
– ChatGPT on its own writes course content. It can’t make a course, because it has nowhere to put one – everything it produces has to be rebuilt by hand in another tool.
– The connected method uses ChatGPT Developer Mode, which turns on custom apps and lets you add an MCP server. Setup is four steps and about a minute.
– Connected, one prompt returns structured lessons with a quiz after each, interactive blocks, and a live URL. Nothing gets pasted.
– The biggest quality gain isn’t the writing. It’s that the assistant can read your source material where it already lives, so the course reflects your business rather than the internet’s average opinion.
– ChatGPT still won’t decide what to teach. A vague brief produces a polished course about nothing, and that hasn’t changed.
Method 1: The Chat-Only Way, and Where It Stops
This method is genuinely worth knowing, because it needs nothing but a ChatGPT account, and for a single short course it’s fine.
Give it your material, not a topic. The difference in output quality between “write a course about customer onboarding” and “here’s our onboarding runbook, build a course from it” is not subtle. The first produces a competent article about onboarding in general. The second produces something about how your company does it.
Ask for structure, then content, in that order. One prompt for a lesson-by-lesson outline with objectives. Fix the outline while it’s cheap to fix. Then a separate prompt per lesson. Asking for everything at once reliably gets you five thin lessons rather than five real ones.
Make it write the quiz separately, against the objective. Quizzes generated in the same breath as the lesson tend to test whether you read the paragraph. Asking for questions that test whether someone can apply the rule to a case you describe produces something worth scoring.
Where it stops. You now have text in a chat window. Everything after this – lessons, formatting, quizzes, interactivity, access, publishing, tracking – is manual, and it’s most of the work. A five-lesson course is maybe two hours of assembly. Then the policy changes in March and you do a chunk of it again.
That ceiling is structural. ChatGPT isn’t connected to anything, so it can’t put the course anywhere.
Method 2: Connect ChatGPT to Somewhere the Course Can Live
The second method removes the assembly step by giving ChatGPT access to a course platform. It’s the same conversation you were already having, except the output is a course instead of a description of one.
The mechanism is MCP, the Model Context Protocol – an open standard that lets an assistant act inside outside systems. If you want the concept without the setup, we wrote a plain-English version for L&D teams. Otherwise, the setup is below.
How to turn on ChatGPT Developer Mode
Custom MCP connections live behind Developer Mode, which is a settings toggle rather than a developer tool despite the name. Four steps:
- Turn on Developer Mode. Enable it in your Apps settings.
- Create a new app. Settings, then Apps, then Create.
- Add the MCP server URL. For Mini Course Generator that’s
https://mcp.minicoursegenerator.com/mcp. - Sign in and use it. Complete the OAuth sign-in, then enable the app from the tools menu in any chat.
That’s the whole setup, and it’s a one-off. The sign-in authorizes the connection against your own account, so ChatGPT acts with your permissions and can’t reach anything you couldn’t reach yourself.
The prompt that actually works
With the connection live, the prompt shape changes. You’re no longer asking for text, you’re asking for a course to exist.
“Using the onboarding runbook in my Drive, build a course in Mini Course Generator. Five lessons, one per stage of the process. Add a quiz after each lesson that tests whether someone could apply the step, not whether they read it. Add a branching scenario for the escalation path in stage four. Publish it and give me the link.”
Three things in that prompt do the heavy lifting:
- “Using the runbook in my Drive” – because ChatGPT is already connected to your document tools, you name the source instead of pasting it. The course is built from your actual material.
- “tests whether someone could apply the step” – the same quiz instruction as method one, and it still matters.
- “Publish it and give me the link” – the instruction that method one physically cannot execute.
What comes back is a live course you open, review and edit. It is not finished work. It’s a first draft that already exists in the right place, which is a different starting position from a chat transcript.
What You Get vs What You Paste
| ChatGPT alone | ChatGPT connected over MCP | |
|---|---|---|
| Lesson content | Text in the chat | Lessons inside a course |
| Quizzes | Questions you rebuild | Attached to lessons, scored automatically |
| Interactive blocks | Described, not built | Flip cards, hotspots, branching scenarios, built |
| Source material | Whatever you pasted | Read from Drive, Notion, Confluence, Slack |
| Publishing | You do it | A URL, from the same prompt |
| Learner tracking | None | Completion and per-question results |
| Updating it in March | Repeat the assembly | Ask for the change |
The row that changes the arithmetic is the last one. First builds are exciting and rare. Maintenance is boring and constant, and it’s where the manual method quietly costs the most – a point we made at length in what an agent does after publish.
Does ChatGPT Create a Full Course or Just an Outline?
This is the question most people actually have, and the answer depends entirely on whether it’s connected.
Unconnected: an outline and lesson text. Very good lesson text, often better than a first human draft. It cannot produce a course, because a course is a structured, deliverable, trackable object and ChatGPT has nowhere to put one.
Connected: a full course. Lessons, quizzes, interactive elements, published, with learners able to take it and results coming back.
The distinction gets blurred in a lot of marketing, so the test is simple: at the end, do you have a URL you can send to somebody, or do you have text you have to do something with?
What ChatGPT Still Won’t Do
Three limits, and they survive the connection.
It won’t decide what to teach. Connected, it builds what you ask for very fast, which means a vague request produces a polished course about nothing in particular. Knowing that new hires stall in week two on one specific escalation path is a human judgement, and it’s the input that decides whether any of this was worth the effort.
It won’t remove review. Content generated from your own documents is far more reliable than content generated from the open internet. It’s still generated, and somebody who knows the subject reads it before learners do – especially anything with a compliance consequence.
It won’t fix a bad source. Building a course from documentation that’s been wrong since last spring produces a well-structured, confidently-worded, wrong course. The connection makes the pipeline faster in both directions.
If you’d rather do this in Claude, the connector flow is different and the Claude walkthrough covers it. If you live in a code editor instead, Cursor, Copilot and Manus connect to the same server. The architecture underneath is identical.
The Short Version
There are two honest answers to how to create an online course with ChatGPT.
The chat-only method works and has a hard ceiling: you get excellent lesson text and then do two hours of assembly per course, again every time something changes. Give it your material rather than a topic, ask for structure before content, and write the quiz against the objective, and you’ll get the most out of it.
The connected method turns on Developer Mode, adds an MCP server, and moves the whole job into the conversation. One prompt reads your existing documentation and returns a published course with quizzes and interactive blocks, and the same conversation handles the update in March.
Neither method decides what to teach. That part is still yours, and it’s still the part that determines whether the course was worth building. If you want the connected version, the MCP server takes about a minute to set up, and the honest test is to point it at a document you’ve been meaning to turn into training for six months.
Frequently Asked Questions
Can ChatGPT create a full online course?
On its own it creates course content – outlines, lesson text, quiz questions – which you then assemble somewhere else. Connected to a course platform through an MCP server, it creates the actual course: structured lessons, attached quizzes, interactive blocks and a published link.
What is ChatGPT Developer Mode and do I need it?
Developer Mode is the setting that lets you add custom apps and MCP servers to ChatGPT. You need it for the connected method. Despite the name it’s a toggle in Apps settings rather than a coding tool, and the setup is four steps.
Is it safe to connect ChatGPT to our course platform?
The connection is authorized with OAuth against your own account, so ChatGPT acts within your existing permissions and can’t reach anything you couldn’t. You can revoke it from the same settings screen at any time. For regulated data, ask any vendor for their compliance position in writing.
How do I stop ChatGPT writing generic course content?
Give it your source material instead of a topic. A course generated from your own runbook, policy or product documentation is about how your company actually works; a course generated from a topic prompt is about the internet’s average understanding of that topic.
Can ChatGPT build the quiz too?
Yes, and the instruction matters more than the model. Ask for questions that test whether a learner can apply the rule to a described situation, rather than recall a sentence from the lesson. Generated in the same breath as the lesson, quizzes default to comprehension checks.
Does this work with the LMS we already use?
If you need the course to run in an existing LMS rather than a new platform, export it as SCORM and import it the way you import any other course. Where you only need one interactive activity inside an existing course, an Agent Skill is the lighter route – it produces a self-contained SCORM package with no account required.
