I’ve been working through how to add a reliable help system for the ham radio practice exams, and I wanted to capture what I’ve learned before it slipped away. The process involved a mix of free attempts, some not-so-free surprises, and ultimately some promising experiments using GPT-5 and canvases. This write-up is both a set of notes for myself and a guidepost for anyone trying to do something similar. Looks like I'll be heading towards the OpenAI soon. I was hoping the help system could be free like the exams, but perhaps, not so much.
What Can Be Done For Free?
As it turns out, not much that's reliable, but some things that are a bit clunky, which, when they work, work really well.
The Unbeknownst to Me Not Free Options
GPT Code Interpreter Running the Practice Exam
This one was a bit of a no-brainer. Asl GPT-5 to convert the JavaScript exam code to Python and then run it in the code interpreter, (because Python is allowed, but JavaScript isn't.) This worked well, but it was slow. I know how much code is executed to construct and display each question. When GPT-5 was offering a full tutorial per question, the delay was still a little annoying, but understandable. Once I modified the Project 'Instruction', (read prompt), to stop the tutorials, the speed was the same.
Plusses:
- GPT-5 taught me how to deploy the project
- The tutorials while slow, were pretty phenomenal. The app even caught an incorrect exam pool answer that has since been removed.
Minuses:
- I fought and fought to find the correct combination to display figures that come with some exam pool questions.
- I fought less, but still did to display the scoring progress bar charts.
- To use Code Interpreter, you have to be on a paid plan.
GPT Canvas Running the JavaScript Exam
A day after I started, Simon Willison published the post about using canvases in GPT. The application that already exists here on the blog cam up seamlessly in a GPT canvas. I suspect the URL state storage is going to be an issue. I did develop a help system that was as good as the version mentioned above. I stopped when I realized that deploying with a Canvas would also require users to have a non-free OpenAI account.
Plusses:
- Even though this isn't what canvases are intended for, (they're supposed to be for building apps that live independently of ChatGPT), the app came up very, very easily in the canvas and worked well with no delays.
Minuses:
- GPT-5 could have made me aware of this when I asked if it could run JavaScript.
- I suppose technically it can't since it can't move data back and forth from the canvas.
- Again, users have to have a pay account to use this sort of project.
- Again, while the saved state idea could have been made to work, I don't believe it would have worked as a URL like the practice exams here.
This method works... sometimes. On my paid account, after extolling the LLM to actually use the most up to date question pool twice, it finally provided excellent help, just like in the first option, for each question. On a free account that I shared the chat to, no matter how many times I asked GPT to be sure to use the latest question pool, it didn't.
And Where is the Help Info Coming From?
Here's where we start to fall down into SolderSmoke's AI Apocalypse. When I asked to see the sources for the tutorials, the frequently referenced KB6NUs exam material as well as some HamExam material. Sigh... At least there were links to direct the user to the unfiltered, unmitigated sources. I don't know quite what to think of all this yet. As I mentioned earlier this week, this blog is occasionally being referenced as well.
In the end, building a ham radio exam help system isn’t just about code or clever hacks—it’s about balancing usability, speed, and accessibility. The experiments with GPT-5 taught me a lot: sometimes the paid options are worth it, sometimes the free ones almost get there, and sometimes the quirks themselves are part of the learning process. I’ll keep refining this approach, and in the meantime, maybe these notes can help someone else chasing the same problem.
Comments
Post a Comment
Please leave your comments on this topic: