The Simulation Builder
I tested three AI tools to build software simulations. None gave me everything I wanted — so I built my own web app that does. Drop in screenshots, place the hotspots exactly where I want them, control the pacing, and export a perfectly reproducible course every single time.
Produce reproducible, standardized software simulations at scale — identical output every time — without depending on any single AI tool staying available or in budget.
Each AI tool could build a one-off simulation, but none guaranteed identical, repeatable results — a blocker for the consistent, maintainable courses enterprise training depends on.
(drop a screen recording into sims/ and swap this slot for a <video> block)
Building a simulation end-to-end in the app — screenshots in, finished SCORM out.
Three tools, one missing piece
I'd just spent a week running the same watch-me / try-me simulation through ChatGPT, Copilot and Claude. Each could do the work — but every one of them shared the same flaw: I couldn't guarantee a reproducible result. Same prompt, slightly different output, every time. For a training team that needs consistent, standardized courses, that's a real problem.
I was going back and forth with ChatGPT about reusability and standardization when it made a suggestion that changed everything: instead of prompting — or even writing a GPT — why not build a React Native Web app that produces the output for me?
ChatGPT wrote the seed. Claude grew it.
It suggested a React Native Web app — and wrote the initial code for me to hand off. The idea: define the output once, drop in screenshots, package it up.
I imported that seed into Claude and kept prompting — shaping it to look and feel like the polished Claude output I'd loved in my testing. Usage limits stretched it across a few days, but the app took shape.
A web app I own outright. I return to Claude only when I want to add a feature — then re-export. No live dependency on any AI tool staying online or in budget.
A look at the builder
Screens of dropping in screenshots, placing hotspots, and exporting the finished course.
e.g. dropping in screenshots
e.g. placing a hotspot
e.g. the exported course
Drop images into the images/ folder and swap each dashed slot for an image — there's a ready-to-use snippet in the page comments.
The best outcome of the whole experiment
It produces the exact same output in the same wrapper, every single time — the one thing none of the three AI tools could promise. That's what makes it usable for standardized, enterprise training.
Once a course is built I don't have to return to Claude — and I don't lose access if an AI tool is shut down, throttled, or pulled from budget. The app is mine.
I adjust exactly where each hotspot lands and control the pacing myself — no more fighting an AI over hotspot placement, the single biggest struggle across all three tools.
I built in export & re-import of working files — so I can hand a project to a colleague, stop and pick it back up later, or reopen a course to edit when an update is needed.
In testing — and already growing
The app is in testing now. I've handed it to a colleague for feedback, and I'll iterate and refine until it does exactly what we need — verifying that every course it builds is reproducible the way we want it.
And we're already planning the next big feature: branching selections, so a single simulation can adapt to the path a learner chooses.
It started as a simple experiment
A week of testing AI tools turned into a tool of my own. See the full comparison that led here, or explore the wider AI-enablement work behind it.