What Mindsmith Changed About Our Learning Pages

We run a training area on one of our directories. Thirteen tracks, each one a single page. The brakes track alone was 13,655 words. On a phone that is 91 screens, and the only way to get anywhere was a row of jump chips at the top.

It worked. Nobody finished anything.

Then somebody sent us Mindsmith, an AI native eLearning authoring tool, with the note "look at this for the learning area." So we spent an afternoon in it. It is a good product, built by people who understand instructional design a great deal better than we do, and that afternoon was worth more than a month of our own guessing. This is what it taught us.

What Mindsmith is doing

The pitch is "create your best eLearning with a prompt." You feed it a document or a deck, an agent turns it into a lesson, and you refine that lesson in an editor built for instructional designers rather than for developers. Thirty or so interactive element types, real time collaboration, SCORM export, seventy languages, an MCP server so other agents can drive it.

The thing that stopped us was not the AI. It was the unit of work.

In Mindsmith a lesson is the thing. Not a course, not a module. A lesson is short, it has one objective, it ends, and you can tell whether somebody finished it. Everything in the product is built around that being true.

Our thirteen pages were not lessons. They were textbooks with a table of contents bolted on.

The split was free, because the data already agreed

Here is the part that mattered and that we nearly missed.

Our lessons live in one JSON file per track, and our exams live in another. When we went to break the tracks apart, we found the topics in the lesson file matched the topics in the exam file one for one, across all thirteen tracks. Every single one.

That is not a coincidence. It means a skill area was already a real unit in our own data, and had been since the day we wrote it. We had been rendering it as a chapter heading instead of as a page.

So a track became a hub. Teaching moved down one level.

• 189 lesson pages where there had been 13
• 214 learn URLs in the sitemap, up from 14
• About 50 KB a page, down from 212 KB

None of that required new content. It required reading our own data and noticing it disagreed with our own layout.

Progress had to survive the move

This is the thing that would have made the rebuild not worth doing.

People had already cleared concepts on the old long pages. If the new pages forgot that, we would have reset every learner on the site to zero to give them a nicer layout. That trade is not worth making and you usually do not find out until somebody complains.

We got lucky, and the luck was old caution. The progress store keys on the text of the concept, never on its position or its page. So a concept cleared on the old 13,655 word page is still cleared on the new short one, because it is the same words.

If you are about to reshape pages that hold any user state, go find out what that state is keyed on before you touch anything. If the answer is an index or a page number, fix that first as its own change.

Work the call

Mindsmith has branching scenarios. We took the idea and gave every lesson one, called Work the call.

It opens with a real situation, then walks through up to three decisions. What it does not do is generate anything. Every option, every consequence, every explanation is human written content that was already in the repo, pulled from the authored checks and the exam questions for that skill area. The component picks and arranges. It never invents.

That rule is worth writing down somewhere permanent. An AI built scenario about brake fluid boiling points that invents a number is not a teaching tool, it is a liability. Arranging content a human wrote is safe. Producing content nobody reviewed is not.

Two densities of the same thing

The second thing we learned there, and it mattered more than we expected.

Mindsmith offers a guided walkthrough, one step at a time, and an expert summary, the whole thing on one page. Same content, two readers. A beginner needs the scaffolding. Somebody with fifteen years on the job wants the page and wants to leave.

We had already split our learners into tiers and then served both of them the same wall of text anyway. Having the shape named made that obvious.

Four bugs only the live page could find

We ran a structural checker over the new pages. Every link resolved, every component rendered, every page built. It passed clean.

Then we opened it on a phone.

1. Concept rows inherited a card style that carried a permanent green tick. Every row read as cleared while the counter underneath said 0 of 10.
2. The navigation rail rendered above the heading on phones, so every lesson opened on a table of contents instead of on the lesson.
3. Fixing that caused the worst one. Moving the rail meant switching the layout to a column flex container, and a column flex container stretches its items to their minimum content width. The main column rendered 462 pixels wide inside a 375 pixel phone and every line of text was clipped at the right edge. The document scroll width still read 375, so an overflow check on the document would have called it clean. Grid with minmax(0, 1fr) is the track that refuses to grow past its container, and grid honors ordering too.
4. The progress line repaints on load with English labels baked in, so the Spanish pages rendered correct Spanish and then overwrote themselves a beat later.

Not one of those is a broken link. A checker proves the pieces exist. It cannot tell you the result reads correctly, and on a page whose entire job is reading, that is the only thing worth knowing.

Credit where it belongs

None of the above is a knock on Mindsmith, and none of it is a substitute for it. If your job is authoring eLearning, the honest answer is to go use the product. It does a long list of things we have no intention of building, including SCORM export, real time collaboration, seventy languages, and an editor somebody who is not a developer can actually work in.

Our problem was narrower than theirs. The content already existed, in our own repository, tied to our own exams, and it had to live inside a directory we run and rank. We were never going to be their customer for this particular job. What we needed was a clearer idea of what good looks like, and an afternoon with a tool made by people who do this for a living gave us that.

So that is the recommendation, and it is the whole point of writing this up. Before you design a kind of thing you have never built before, go and spend real time inside the best example of it you can find. You are not there for their code. You are there to find out which decisions are load bearing, because that is the part that is genuinely hard to work out from a blank file.