The honest version of how to create a training module is that you almost never build one from nothing. You build it from a policy legal wrote, a deck someone presented three years ago, an SOP sitting in a shared drive, or a recording of a session that already happened. That makes the job conversion, not authoring, and conversion is where AI drafting earns its keep and where it fails in two specific places you can point at before anything gets assigned. This guide walks the five things you’re likely starting from, what to hand the model and in what scope, and the checks that have to happen before the module goes out to a few hundred people and into an audit file.
A training module is the unit that gets assigned, completed and recorded: one thing a person has to be able to do differently, with one completion record attached to it. A training program is a sequence of those. Hold that distinction and most of the sizing questions answer themselves.
Key Takeaways
- Half the legally required content of a hazard communication module can’t come from a model. Two of the four training requirements describe your building, your equipment and your labeling system.
- The refresh trigger for that module is an event, not a date. The word “annual” doesn’t appear in the standard’s training paragraph.
- 10 minutes of reading is roughly 2,400 words, so an hour-long recording is a cutting job at a ratio of several to one.
- Scope the source before you upload it. The six to ten sections that carry the decision beat the whole handbook, and they beat five regenerations of the whole handbook.
Convert documents into courses – the deck and the policy you already have are the raw material, not a blank editor.
What a Training Module Is, and Why a Policy Is Not One
Take the training paragraph of the federal hazard communication standard, 29 CFR 1910.1200(h). It runs 300 words, averages 37.5 words per sentence, and scores 5.5 on the Flesch reading ease scale. Its training subparagraph, (h)(3), is a single sentence of about 170 words carrying four separate requirements.
That document is the source for a module thousands of companies have to run. It is also unreadable by design, because it was written to be enforceable, not followed. The rewrite isn’t a step on the way to the module. The rewrite is the module.
Which is why “generate a module from this” is the wrong mental model to start with. The work is splitting one 170-word sentence into four things a person on a loading dock can do, in words they use, and then deciding which four and whose words. A model is good at the splitting. It has no opinion on the deciding, and the deciding is what you’re paid for.
The sizing rule that follows is short enough to apply today. One module equals one thing a person has to be able to do differently, which is also one completion record. If you can’t write that one thing in a sentence a supervisor would recognize, you have a program on your hands and it needs breaking up before anything gets drafted. Where a single module sits inside the wider picture is the subject of AI in learning and development.
Start From What You Already Have: Five Source Types
The five things you’re most likely starting from are not equally usable, and each has one step that has to happen before the model sees it. Skip the step and you get a draft that reads well and is wrong in a way that only shows up when somebody on the floor follows it.
| Source | What it gives you free | What it’s missing | The step before the AI sees it |
|---|---|---|---|
| Policy or regulation | Objectives, scope, the non-negotiables | Anything about your building | Cut to the paragraphs that bind this audience |
| Slide deck | Structure, sequence, visuals | The narration, which was the teaching | The deck test below, or 15 minutes with the presenter |
| Recorded session | Real language, real examples, real questions | Editing. It’s several times too long | Export the transcript and read its length |
| SOP or runbook | Steps in order, already correct | Why, and what happens when it goes wrong | Decide which decision the learner has to make |
| Help center or wiki | Current, maintained, already written down | A sequence. It’s built for lookup, not learning | Pick the six to ten pages that carry the decision |
The bottom two rows are the ones most teams have most of, and they fail in opposite directions. An SOP in a Google Doc is already correct and already in order, which is why it converts so cleanly and teaches so badly. It tells someone what to do and never why, so the first thing you add by hand is the failure case: what goes wrong when step four gets skipped, and who pays for it.
A Confluence space, a Notion page or a set of help center articles has the reverse problem. The content is current and someone else maintains it. But it was written for someone who already knows what they’re looking for, so it has no sequence at all. You supply the sequence, and the model fills it in.
The deck test
Open the deck and read it as if you’d never seen it presented. The test isn’t a word count. It’s a question: can a person who wasn’t in the room tell what they’re supposed to do from the slide alone?
If the slides are titles, bullets and a chart, the teaching was the narration and the narration was never recorded. Hand that deck to a model and you get headings joined by invented connective tissue, and the invented tissue is where the errors are. It reads fluently, which is what makes it dangerous.
The fix isn’t a better prompt. It’s 15 minutes with whoever presented it, recorded, or a different source entirely. A deck that passes the test converts into e-learning as cleanly as anything in this category. A deck that fails it is a prompt for a conversation, not a source.
A recording is a source with a length problem
The average adult silent reading rate for English non-fiction is 238 words per minute, from a meta-analysis of 190 studies and 18,573 people, with most adults falling between 175 and 300. So 10 minutes of reading works out at about 2,400 words, and that number is the one firm constraint you have when you’re staring at a transcript.
An hour-long session transcribes to several times that, so export the transcript and read its length instead of estimating from the meeting length.
The model’s job on a recording isn’t writing, then. It’s throwing away, and most of that transcript goes. Deciding what gets thrown away takes someone who knows which two minutes of that all-hands changed how the work is done, and that person is you, not the tool.
Scope Before You Upload
The order that works is the reverse of what every tool’s onboarding suggests. Four steps, in this order:
- Name the one decision the learner has to make on the job.
- Find the six to ten pages or sections of the source that contain that decision.
- Cut to them. Delete the rest from the file you’re about to upload.
- Then generate.
Hand over a whole space or a whole handbook and the draft comes back thin in the middle, with the parts that matter flattened into the trivial ones. Every section gets roughly equal weight, because nothing in the source told the model which one a person could get wrong at real cost. That’s not a prompt failure you can fix downstream. It’s a scoping failure you built in at upload.
Every minute spent scoping saves more than it costs, and it’s the difference between a module and a summary. It’s also why re-prompting is the wrong reflex. “Regenerate until it looks right” is a loop with no convergence, and a clean, scoped source produces a better first draft than five regenerations of a bad one.
Where Your Objectives Come From, and When They Are Already Written
For a regulated topic, your learning objectives have already been written, by the regulator, and copying them out is the correct move, not a shortcut. The hazard communication standard is the clearest case because it enumerates them.
Paragraph (h)(2) lists three things employees have to be informed of: the requirements of the standard itself, any operations in their work area where hazardous chemicals are present, and the location and availability of the written hazard communication program, including the required chemical lists and the safety data sheets. Paragraph (h)(3) then lists four things training has to include at minimum: how to detect the presence or release of a hazardous chemical, the hazards of the chemicals in the work area, the measures employees can take to protect themselves, and the details of the employer’s own hazard communication program.
That’s seven numbered, non-negotiable items. They’re your objectives. Write each one as a plain sentence, build the module against the list, then check the finished module back against the list before it goes out.
Never let a model invent objectives for a regulated topic. It’ll write plausible ones, and plausible isn’t the standard an auditor applies. If compliance is most of what you own, compliance training software is the surface this work lives on.
The second thing the rule settles is the refresh cadence, and most teams get it backwards. The trigger in (h)(1) is training “at the time of their initial assignment, and whenever a new chemical hazard the employees have not previously been trained about is introduced into their work area.” The word “annual” doesn’t appear anywhere in the paragraph.
That changes the build. The refresh trigger is an event, not a date, which argues for small, per-hazard modules you can re-assign one at a time instead of one annual course everyone re-takes. Record the trigger where you record everything else about who owes what, which is your training matrix. Plenty of teams run an annual cycle they were never required to run and miss the new-hazard trigger they were.
What AI Drafts Well, and the Two Paragraphs It Cannot Write
Go back to the four requirements in (h)(3) and sort them by who can possibly know the answer. Items (i) and (ii) are about detection methods and about the classes of hazard present, and both can be drafted from the published rule and a chemical list. A model handles them well, and that’s real time saved.
Items (iii) and (iv) are a different kind of thing. Item (iii) requires training on “specific procedures the employer has implemented to protect employees from exposure to hazardous chemicals, such as appropriate work practices, emergency procedures, and personal protective equipment to be used.” Item (iv) requires “the details of the hazard communication program developed by the employer, including an explanation of the labels received on shipped containers and the workplace labeling system used by their employer.”
Both exist only in your building, on your shift, with your equipment and your labeling system. Both are legally required. So half the required training content in this module cannot come from the model, and if the model produced text for (iii) or (iv), it invented it. Those two paragraphs are yours to write, and they’re the two an auditor reads most closely.
That’s the general shape, not a quirk of one standard. Every module has a part that’s true everywhere and a part that’s true only here, and the second part is the reason the module exists.
What goes in, and what doesn’t
Be honest about the doors. In Mini Course Generator, existing material gets in through exactly three of them:
- PDF to Course Creator, for anything you can print to PDF.
- AI YouTube-to-Course Builder, for a video that’s already on YouTube.
- Text pasted into the editor, for everything else.
The AI Lesson Page Generator and the AI Quiz Builder generate content once you’re inside. They aren’t ingestion paths, and treating them as ways in wastes an afternoon.
Which means a recorded Teams all-hands has a step before the tool: get a transcript, or produce a PDF. That’s a five-minute job if you know it’s coming and a frustrating one if you find out at upload.
Fidelity is not truth
Mini Course Generator’s Hallucination-Proof is described on its own page as “a ‘fact-checking assistant’ that measures how well a rewrite maintains the factual integrity of the original text,” producing “a numerical score that tells you how well the generated text preserved the facts from the original.”
That’s fidelity to your document, and the precision is the useful part. It catches the model inventing a threshold your policy never contained. It can’t catch your three-year-old deck being wrong about the current process, because the deck is the ground truth it’s comparing against.
So there are two checks and only one of them is automated. The other is you, reading the module once with the source open, looking at nothing but the numbers. Thresholds, dates, dollar figures, deadlines and quantities are what a paraphrase silently moves, and they’re the ones that hurt.
Quiz Questions That Test the Job, Not the Sentence
A generated quiz tests whether someone read the module. Three small moves turn it into one that tests whether they could do the job, and together they cost about five minutes.
Ask for three options, not four
A meta-analysis covering 80 years of measurement research found three options optimal for multiple-choice items (Rodriguez, 2005). The fourth option is nearly always filler nobody would pick, and filler makes a question easier without making it fairer.
One word in the prompt fixes it for good. Ask for three options and the model writes three real ones instead of two and a joke.
Three defects to check in 30 seconds
Three defects show up in generated items often enough to be worth a pass of their own. Read the quiz looking only for these.
- “All of the above” appearing as an option. It rewards partial knowledge: someone who is sure of two of the three can pick it without knowing the third.
- A clue to the answer sitting in the stem, most commonly a word repeated between the question and the right option.
- One option noticeably longer than the others. The longest option is usually the right one, and learners work that out by question three.
That’s a five-minute read for a quiz of 10 questions, and it decides whether the completion score means anything to the person who asks for it.
Ask for questions from the mistakes, not the content
This one’s a rule of thumb, not a finding, and it’s the highest-return change you can make to a prompt. “Write five questions on this module” produces questions about sentences. “Here are the three mistakes people make doing this job, write five questions about those” produces questions about decisions.
The list of mistakes is the input only you have. It takes two minutes to write down and it changes what the quiz is for. If you work with an instructional designer, this is the point where their judgment pays for itself, and AI for instructional design goes further into where that judgment sits in an AI-drafted workflow.
An Example Module, Filled In
Here’s the whole thing at your scale, as an example, not a template with brackets in it. It’s built from the hazard communication requirements above, so you can see exactly which parts a model can draft.
- Title – Hazard Communication: Before You Open a Container
- Audience – warehouse and production staff, at initial assignment
- Length – about 10 minutes of reading, so roughly 2,400 words
- Refresh trigger – a new chemical hazard entering the work area. Not a date.
Objectives, each one rewritten as a sentence a supervisor would recognize:
| # | The learner can | Who supplies the content |
|---|---|---|
| 1 | Say what this standard requires of them and of the company | The rule |
| 2 | Name the operations on their shift where hazardous chemicals are present | The rule, plus your list |
| 3 | Find the written program, the chemical list and any safety data sheet | The rule, plus your locations |
| 4 | Detect the presence or release of a hazardous chemical by sight, smell or monitor | The rule |
| 5 | Name the hazard classes present in their work area and what each one does | The rule |
| 6 | Follow the work practices, PPE rules and emergency procedure used here | You. If the model wrote it, it invented it. |
| 7 | Read a shipped container label and this site’s own labeling system | You. If the model wrote it, it invented it. |
Sections, six, each one screen, each named for what the person does:
- Before you open a container: what you look at first
- Reading the label on a drum that just arrived
- Finding the safety data sheet for what’s in your hand
- What’s on this site, and what each one does to you
- What we do here: work practices, PPE and the spill procedure
- When something’s wrong: who you tell, and where the eyewash is
The one question, drawn from a mistake rather than a sentence:
A drum arrives on the dock with the shipping label torn half off. The product name is readable, the pictograms aren’t. What do you do before it goes on the rack?
a) Rack it and tell your supervisor at the end of the shift
b) Stop, leave it where it is, and get your supervisor now
c) Look the product up, print a replacement label and apply it yourself
The completion evidence line carries learner name, date completed, module version, and the answer given on the question above, held wherever your audit is run from.
Copy the shape and fill it with your own hazard, policy or SOP. The part that takes judgment is the last column of that table, and it’s the same column on every module you’ll build.
Before You Assign It: Accessibility, the Record, and Who Owns It in March
Five checks stand between a draft and an assignment. Four of them are yours. This is the whole list.
| The check | What it catches | Who does it |
|---|---|---|
| The paragraphs that describe your site | A model inventing your procedures and your labeling system | You, first |
| The numbers, with the source open | Thresholds, dates and quantities a paraphrase moved | You |
| Alt text on every image | Captions that describe the picture and not the teaching point | You |
| Where the completion record lands | An audit answer split across two systems | You, with whoever runs the audit |
| Is it still true? | A source that went out of date before the module shipped | The subject matter expert |
The first two are covered above. Here are the three that need their own note, in the order you hit them.
Alt text that describes the point
Auto-generated alt text describes the picture. Instructional alt text describes the point. A photograph captioned “a worker in a warehouse” fails completely when the teaching point is that the drum behind him is unlabeled.
One question per image, and it takes seconds: does the alt text say what the learner is supposed to notice? It matters more again if your modules fall under a public-sector accessibility requirement, and the accessibility guide for course creators covers the rest of the pass.
Decide where the completion record lands
If the module is authored in one system and the record of who finished it lives in another, the auditor’s question has two half-answers instead of one. Show me who did it and when should return one list from one place.
The rule takes a minute to apply.
- If the audit is run out of the corporate LMS, export the module and let the LMS hold the record.
- If there’s no corporate LMS, keep the record where the course lives, and stop pretending otherwise.
Both arrangements work, which is why this is a decision and not a default. xAPI and SCORM compares the mechanics of each. And if the module ends in something the learner keeps, decide now how you’ll issue a certificate with the completion record instead of bolting it on later.
A standard SCORM package is a frozen copy
Export, upload, and the course sitting in the LMS stops tracking the source. It’s a copy, and it stays exactly as it was on the day it went in.
So when the policy changes in March, you re-export and re-upload everywhere it went, which means keeping a list of everywhere it went. That cost belongs in the build decision, not in next year’s surprise, and it’s the strongest argument for keeping one canonical version and exporting from it. If you’re a training provider building into a client’s LMS, settle this in writing before the first module ships, because the re-export is yours to do and rarely yours to bill.
Check the System You Already Pay For
Before anyone starts a procurement cycle, open the admin side of your own HRIS or LMS and look at what it does now. The category moved fast and the systems already on your invoice moved with it.
Workday Learning’s own product page describes built-in authoring with an AI writing assistant, “removing the need for heavy production cycles.” It names an AI Authoring Agent that “[e]nable[s] subject matter experts to turn raw knowledge into world-class content in seconds,” and course creator agents for creating, updating and translating content.
So the first question for a lot of teams isn’t which tool to buy. It’s whether the one already on the invoice does what this particular module needs. If it does, the module is a two-day job instead of a quarter-long purchase, and you’ve spent ten minutes to find out.
If the answer is no, you’re buying for the conversion job rather than for a course catalog, and that narrows the shortlist considerably.
Mini Course Generator
Mini Course Generator is our product, so treat this section as interested rather than neutral and check it against the pricing page before you act on it.
It’s built for the conversion this article describes rather than for authoring from a blank page. You hand it the PDF, the YouTube link or the pasted text — the three doors above — and it comes back as a course with pages and a quiz you can edit. The first draft arrives in the shape a learner moves through, instead of as a wall of text you then have to lay out.
Two things line up with the checks above. Hallucination-Proof scores how well the generated text preserved the facts in the source you gave it, which narrows your review to the paragraphs you always have to write yourself: your building, your equipment, your labeling. And the accessibility widget and voice-over sit in the product rather than in something you have to remember to bolt on, though the alt-text pass in the checklist above is still yours to do image by image.
On the record question you can go either way without rebuilding. Courses hold their own completion records and analytics when there’s no corporate LMS behind them, or export to whichever system runs the audit. PDF export and SCORM export both exist, and there is a separate dynamic SCORM export as well. Which tier carries which is on the pricing page, and that is the page to check rather than this one.
Ownership is settled in writing rather than left to be inferred. The pricing FAQ puts it plainly: “Your content is always yours, whatever your subscription status. You can export mini-courses as PDF files or SCORM packages at any time.”
Pros. Turning a document into an assignable module is what the product is organized around, so the conversion path is the front door rather than a submenu. Badge removal is included on every paid plan rather than held back for a higher tier. And every plan opens with 14 days of full access and no card, which is enough to run one real policy through it end to end.
Cons. The format is deliberately short, so a multi-day curriculum with dozens of modules isn’t what it’s shaped for. There’s no permanently zero-cost route, so if the budget is genuinely nothing, this isn’t the section for you. Learner allowances are counted per year, which means a company-wide compliance push can consume the year’s allowance in one week. And learner SSO and the REST API sit at the top of the ladder, which matters if the module has to run inside an app you already own.
Plans and pricing lists what each tier includes. The honest test is to put one real policy through it and hold the draft against the module you’d have written by hand.
The Short Version
Open the source you already have and run one test on it. For a deck, that’s whether someone who wasn’t in the room could act on the slide alone. For a recording, it’s the transcript length against the 2,400 words that fill 10 minutes of reading. For a policy, a wiki or an SOP, it’s cutting to the six to ten sections that carry the decision the learner has to make.
Then generate, and then write the parts only you can write: what happens in your building, on your shift, with your equipment. Check the numbers against the source with both open, read the quiz for the three defects, and settle where the completion record lands before you assign anything. That sequence is what separates a module you can evidence from a document with a completion percentage attached to it.
Roll it out with employee training software – assign it, track it, and keep one canonical version you can update in March.
Frequently Asked Questions
What is a training module?
A training module is a single unit of training that gets assigned to a person, completed by them, and recorded against their name. It covers one thing they have to be able to do differently. A training program is a sequence of modules, which is why sizing comes down to asking whether you have one outcome or more than one.
What should be included in a training module?
Five things, and everything else is optional:
- The objectives, taken from the source and not invented.
- The decision the learner has to make on the job.
- A practice or check that tests that decision.
- The completion record, with enough fields to answer an audit.
- The refresh trigger, whether that’s a date or an event.
Can I turn a recorded meeting into a training module?
Yes, with a transcript first, and expect to cut hard. An hour of talking is several times the length of a 10-minute module, so the work is deciding which few minutes changed how the job is done. Get a transcript or a document out of the recording before you reach for a tool, because a raw video file isn’t a usable input on its own.
How do I create a training module for free?
You can, but check what “free” covers before you spend a weekend on it. A free plan and a free trial are different things, and the limits that decide whether either one works for you are the same three every time: how many courses you can keep, how many learners can take them, and whether the tool’s own badge stays on the finished module. The routes that hold up are compared in the guide to create a course for free.
Who has to review it before I assign it?
You do, on four things: the paragraphs that describe your own site, which no model can write, the numbers read once with the source open, the alt text on every image, and where the completion record is going to live. A subject matter expert reviews the fifth, which is whether the module is still true.
Sources
- 29 CFR 1910.1200, Hazard Communication, Electronic Code of Federal Regulations
- Rodriguez, M. C. (2005), Three Options Are Optimal for Multiple-Choice Items: A Meta-Analysis of 80 Years of Research, Educational Measurement: Issues and Practice, doi:10.1111/j.1745-3992.2005.00006.x
- Haladyna, T. M., Downing, S. M., & Rodriguez, M. C. (2002), A Review of Multiple-Choice Item-Writing Guidelines for Classroom Assessment, Applied Measurement in Education, doi:10.1207/S15324818AME1503_5
- Brysbaert, M. (2019), How many words do we read per minute? A review and meta-analysis of reading rate, Journal of Memory and Language, doi:10.1016/j.jml.2019.104047
- Workday Learning, product page
Regulation text is quoted from the Electronic Code of Federal Regulations, and product capabilities from each vendor’s own pages.
