Building a course by hand, on purpose
For the last few months, almost every course I made started the same way: upload a document, pick an instructional approach, wait, refine. It works, and I am not going to pretend otherwise. Then this week I built a course the slow way — from an empty lesson, one block at a time — and I noticed how much I had stopped doing.
Not stopped doing badly. Stopped doing at all.
What the generated workflow gets right
The blank lesson is the hardest part of a course. Not the writing, not the questions — the structure. Where does this start, what are the lessons even called, what belongs in lesson three versus lesson seven. An experienced designer can do this, but it still takes a focused hour or two of staring before the shape appears.
AI course generation hands you that shape. Upload a forty-page policy document, choose whether you want an action-first or objective-first structure, and you get a complete scaffold with explanatory content, knowledge checks, and activities already sequenced. When the deadline is Friday and the source is a dense, well-organised document, this is not a shortcut. It is the only realistic way to ship on time, and I have shipped courses I am happy with this way.
So this is not a piece about AI being bad at course-making. It is a piece about what the fast path skips over, and why that is sometimes worth slowing down for.
The loop that generation skips
When you build a course by hand, you spend most of your time inside a loop. You add an interaction. You look at it. It is not broken — it just is not right. You delete it. You try a different way into the same idea. You sit with a question for a while, wondering whether it is actually testing anything or just checking that someone scrolled past the previous screen. You move a block, dislike it, move it back.
That loop has a lot of dead time in it, and the dead time is where the design decisions actually happen. Trying something and rejecting it is not wasted motion — it is how you find out what you think. By the time a hand-built lesson is finished, you have seen and discarded five versions of it, and the one that survived is better for having competition.
A generated course arrives without that loop ever having run. The decisions were still made — they were just made statistically, by the model, in one pass, and you never got to see the alternatives it discarded. Often the result is fine. But "fine" is exactly the word, and you cannot always tell whether a generated lesson is the best version of itself or simply the most probable one.
Humour and specifics don't survive averaging
The thing I missed most is small: the jokes.
A joke in a course works because it is specific. It lands because it refers to this audience, this office, this piece of software everyone secretly hates, this running gag the team will recognise. AI-generated humour tends to be the average joke — safe, mildly amusing, equally applicable to any audience anywhere, which is another way of saying it connects with no one in particular.
The same is true of examples. The example that makes an abstract point finally click is usually the oddly specific one — the slightly embarrassing scenario, the thing that went wrong last Tuesday, the detail too particular for a model to have reached for. A model reaches for the representative example, because that is what it is built to do. A person reaches for the memorable one. Those are not the same example, and a learner can feel the difference even if they could not name it.
This is an observation, not a research finding. But after building a course by hand, I think the human touch in learning content is mostly this: a steady stream of small, specific choices that no average would ever produce.
Editing a generated course is real work
The other thing I had half-forgotten is how much work it is to amend a generated course into the one you actually wanted.
The demo always shows generation as the whole task — document in, course out. In practice the generated draft is the start of the task. Sometimes reshaping it takes nearly as long as building from scratch: rewriting questions that test recall when they should test judgement, resequencing lessons, cutting a section the model was confident belonged. And it is a particular kind of slow, because you are also reverse-engineering decisions you did not make and quietly arguing with them.
Building by hand, there is nothing to undo. Every decision is yours from the first block, so the work only ever moves forward. Neither path is free. They are just different shapes of work, and it is worth knowing which shape you are signing up for before you start.
I'm not romanticising the slow way
The slow loop is also where procrastination hides. Not every interaction I deleted and retried this week was design judgement — some of it was me fiddling because fiddling is pleasant, and a deadline was comfortably far away. The craft is enjoyable, and enjoyable work is easy to mistake for important work.
Manual authoring is genuinely slower, and for a great deal of content the craft does not matter. A compliance refresher covering a regulation that changed one clause does not need my jokes or my oddly specific examples. It needs to be accurate, clear, and finished. If I hand-built every course, most of them would ship late and most would be no better for the extra time. Pleasure is not the same as value, and confusing the two is how a project quietly slips a month.
So the honest position is not "build by hand." It is "know which courses deserve it."
The point is the dial, not the verdict
LearnBuilder was built so you do not have to settle this argument once and for all. You can generate a full course from a document. You can open an empty lesson and build every block yourself. Or — most usefully — you can generate the scaffold and then spend your craft only where it earns its keep: rewrite the three questions the course actually turns on, hand-make the one branching scenario that matters, add the joke that this specific audience will get, and leave the rest as generated.
The dial moves per course. Honestly, it moves per mood and per deadline too. This week I wanted the slow loop, so I took it, and the course is better for it. Next week, with a policy document and a Friday, I will generate the whole thing and feel no guilt at all. A tool that only does one of those is making that choice on your behalf — and it is a choice that should stay with the person who knows the audience, the subject, and how much time they actually have.
If you want to feel the difference yourself, the free trial lets you generate a course and build one by hand in the same afternoon.