An AI course builder is best understood as a fast, tireless, slightly overconfident junior instructional designer. Give it a syllabus PDF and it will hand back a course outline in a few minutes. Whether that outline is worth teaching depends almost entirely on what you gave it and what you do next.
This post is a run through of one real workflow: taking a syllabus PDF and getting to a first lesson you would actually put in front of learners. What to upload, how to prompt for structure, what the AI reliably gets right, what it reliably gets wrong, the checklist a trainer should run before publishing, and the places to close the tool and write it yourself.
The short version: the AI does the scaffolding and the first draft. The judgement stays with you, and that is not a limitation to work around. It is the design.
Step 1: Get the input right
Output quality tracks input quality more than it tracks prompting cleverness. Before uploading anything, collect the following.
- The syllabus itself, as a text PDF rather than a scan. A scanned photocopy gives the tool a picture of a page; run it through OCR first if that is all you have.
- Who the learners are: role, prior knowledge, language comfort, whether they will study on a phone.
- Time available: total hours, session length, and whether sessions are live, self-paced or both.
- The assessment that already exists, if there is one. Working backwards from the exam keeps the course honest.
- Two or three reference documents you trust: your own standard operating procedure, a manual, a machine handbook. This is what keeps the draft anchored to your organisation rather than to the internet in general.
- Anything that must not be taught: outdated procedures, superseded rules, a competitor’s terminology.
Strip the syllabus down before uploading. University syllabus PDFs carry twenty pages of examination regulations and attendance rules around four pages of actual content. Upload the four.
Step 2: Prompt for structure before content
The common mistake is to ask for lessons immediately. Ask for the skeleton first, correct the skeleton, and only then ask for prose. Fixing a wrong structure after twelve lessons are written costs far more than fixing it at the outline.
- Ask for a module map. Modules, lessons per module, and one line on what each lesson covers. Nothing else.
- Ask for learning outcomes per module, written as things a learner can do, and insist on observable verbs rather than “understand” and “be aware of”.
- Give the constraints explicitly. Number of hours, lesson length in minutes, reading level, language, and the fact that learners are on phones if they are.
- Correct the map by hand. Merge, split, reorder, delete. Expect to change a third of it.
- Now ask for one lesson, the hardest one, not the easiest. The first lesson of a course is always the one the tool writes best, which tells you nothing.
- Ask for the assessment with the lesson, so questions come from the outcomes rather than being bolted on later.
- Repeat with the remaining lessons only once you are satisfied with the format of that hard one.
Useful things to say in the prompt: name the audience and their prior knowledge, ask for examples from your industry and region, ask for short paragraphs, and ask it to mark anywhere it is uncertain instead of filling the gap. That last instruction is worth more than any other single line.
Step 3: Know what an AI course builder gets right and wrong
| Task | Typical quality | What you do about it |
|---|---|---|
| Module and lesson structure | Good starting point, sometimes over-neat | Reorder by what learners find hard, not by textbook order |
| Learning outcomes | Good, occasionally vague | Replace soft verbs with observable ones |
| Explaining a settled concept | Usually clear and well pitched | Light edit for house terminology |
| Summaries and recaps | Reliable | Accept with a read-through |
| Multiple-choice questions | Fluent, but distractors are often obviously wrong | Rewrite distractors around real misconceptions |
| Worked numerical examples | Arithmetic and units slip | Recompute every one by hand |
| Legal or regulatory detail | Unreliable, often out of date | Replace with text checked against the current notification |
| Your internal procedure | Invented unless you supplied it | Write it yourself or paste it in |
| Local and sector examples | Generic, often foreign | Substitute your own site, plant or branch |
| Citations and references | Frequently wrong or non-existent | Open every link before it goes in |
The pattern is consistent. The tool is strong on form and weak on fact, and it is most confident exactly where it should be least trusted. UNESCO’s guidance on generative AI in education notes that these systems remain not fully reliable and can produce facts that are simply invented.
Step 4: The trainer review checklist
Run this before a single learner sees the draft. It takes about an hour for a lesson and it is the step that makes the whole approach defensible.
- Every factual claim traced to a source you opened, or removed.
- Every number, rate, limit or deadline checked against the current official document.
- Every statute or standard named, verified as current, with the date checked.
- Every calculation redone by hand.
- Every citation and link opened.
- Procedures matched against your own written procedure, line by line.
- Examples replaced with ones from your own operations.
- Terminology matched to what your people actually say on the floor.
- Safety-critical content read and signed off by the person accountable for safety.
- Assessment questions checked so each one maps to a stated outcome and the key is correct.
- Language and reading level checked against the actual audience.
- A named reviewer and a review date recorded against the lesson.
That last item matters more than it looks. When a regulator, a customer auditor or a parent asks who approved this content, “the AI drafted it and the safety officer reviewed it on 4 September” is an answer. “It was generated” is not.
Step 5: Know where to stop
Some content should not be drafted by a tool at all, because the review costs more than writing it would have.
- Anything safety-critical. Permit to work, lock-out procedures, emergency response, machine-specific instructions.
- Current statutory detail. Rates, thresholds and deadlines change, and a confidently wrong figure in a compliance course is worse than no course.
- Your own processes. The tool has never seen your dispatch flow or your approval matrix.
- Anything involving real people or incidents. Case studies from your own operations should be written by whoever was there.
- The final assessment. Use AI to draft the bank, but the paper that decides whether someone is certified deserves a human author.
UNESCO’s guidance is direct on the principle: AI must not usurp human intelligence, and human accountability should not be ceded to these systems when high-stakes decisions are being made. Its 2024 AI competency framework for teachers makes the same point from the other side, that AI tools should complement, not replace, the roles and responsibilities of teachers. Treat a generated lesson as a draft from a capable assistant who has never met your learners.
Frequently asked questions
How long does one lesson actually take this way?
Drafting is minutes. Review is the work. For a straightforward lesson, expect roughly an hour of trainer time; for anything regulatory or safety-related, considerably more. The saving is real but it is in the blank page, not in the checking.
Can we put AI-drafted content straight in front of learners?
No, and it is worth writing that into policy rather than leaving it to individual judgement. Publish only what a named person has reviewed, and record the name and date on the lesson.
Do we have to tell learners the content was AI-assisted?
Saying so costs nothing and builds trust, particularly with adult learners who will spot the style anyway. Pair it with the reviewer’s name so the disclosure reads as confidence rather than a disclaimer.
Is there an age issue with learners using these tools directly?
There can be. UNESCO’s guidance strongly recommends age restrictions for general-purpose AI tools and proposes a minimum threshold of 13 years. If your learners are school age, keep the tool on the trainer’s side of the course rather than the learner’s.
Drafting inside the system you teach in
The workflow is easier when drafting, review and publishing live in one place instead of moving text between a chat window and the course. The Quipu LMS AI course builder drafts lessons and summaries from a topic, a syllabus or a PDF for a trainer to review, alongside question banks and the assessment engine. We have also written about what AI-generated course content can and cannot do.
