Building e-learning with AI

Start with the gap, not the tool

The fastest way to waste AI on a course is to point it at the wrong course. Before you open anything, answer one question: what are people doing now that they should be doing differently?

If you cannot name the behaviour, you do not have a training problem yet. You may have a process problem, a tooling problem, or a manager problem, and none of those are fixed by a module. AI will happily build you a beautiful course for a problem that training cannot solve.

• Name the behaviour, the person, and the moment it happens.
• Ask what is stopping them. Knowledge is only one of the possible answers.
• Decide what you will measure before you decide what you will build.

That prompt is doing something specific. It is not asking AI for an answer. It is asking it to interrogate you, which is the part most of us skip when we are busy.

What AI is good at here, and what it is not

Being precise about this saves you more time than any prompt trick. AI is strong on form and weak on truth.

It is genuinely good at first drafts, at restructuring something you already wrote, at generating twenty variations of a scenario when you are out of ideas, at turning a rambling subject matter expert transcript into an outline, and at catching where your own writing goes vague.

It is unreliable on anything specific to your organisation: your policy numbers, your system's actual screens, your regulator's actual wording, what your people actually do. It will produce a confident, plausible, wrong version of all of it.

A useful test before you rely on an output: could this sentence be checked by someone in your building? If yes, have them check it. If no, you probably should not be teaching it.

From outline to storyboard

This is where AI earns its place. You have an outline and a pile of source material, and you need a storyboard: screen by screen, what the learner sees, what they do, what happens when they get it wrong.

Work in passes rather than asking for the whole thing at once. A single giant request produces something shaped like a storyboard that falls apart when you read it closely.

• Pass one: the spine. Objectives, sequence, and where practice sits.
• Pass two: one section fully written, so you can correct the voice and the depth before it is repeated forty times.
• Pass three: the rest, with the corrected section handed back as the pattern to follow.

That last sentence is the important one. Without it you get padding, and padding is the tell that gives AI-built courses away.

Prompts that produce usable course content

Most disappointing output comes from a prompt that did not say enough. Four things, every time.

• Who the learner is, including what they already know. "New hires" and "twenty-year veterans" produce completely different drafts.
• What they must be able to do afterwards, phrased as a behaviour, not a topic.
• The constraint: length, format, tone, reading level, how long the module can run.
• What you do not want. This is the one people leave out, and it does more work than the rest.

Then give it an example of good. A paragraph you wrote, a module you like, anything that shows the target rather than describing it. Showing beats describing every time.

Notice there is nothing clever in it. It is just specific. Specificity is the whole technique.

Media without a studio

Voice, images, and video are where AI has changed the economics of course building most sharply. A narrated module used to mean booking a voice artist and re-recording every time a line changed. Now a script change is a script change.

Three cautions worth holding on to.

• Synthetic voice still needs a script written for the ear. Short sentences. One idea each. Read it aloud yourself first; if you stumble, so will the synthesis.
• Generated images are decoration until proven otherwise. If the image has to show your actual software, your actual equipment, or your actual safety gear, generate nothing and photograph something.
• Say when media is synthetic if your learners would reasonably want to know, and check what your organisation requires. This is moving ground and the polite default is disclosure.

And whatever you generate, caption it. Which is the next lesson.

Accessibility, review, and shipping

Accessibility is not a pass you do at the end. It is a set of decisions you make while building, and retrofitting is roughly three times the work of doing it as you go.

• Alt text on anything that carries meaning. AI drafts this well, and you correct it, because it cannot see what the image is for in your course.
• Captions on all audio and video. Generated captions need a human pass, especially for names, acronyms, and technical terms.
• Colour contrast and keyboard access checked while you are laying out, not after.
• Do not rely on colour alone to signal right and wrong in a question.

Then ship it, and plan the revisit. A course nobody updates is a course quietly going out of date, and AI has made updating cheap enough that there is no longer an excuse.