⟨m∀⟩

The PoC

We started with free ChatGPT 4o account and had the first version of it running under 2 hours.

My son doesn’t have yet the knowledge to drive the AI into the more technical concerns. For instance, the AI decided to store the list of trips in a JSON file, and load it and save it for every command. My son could not judge this as foundational bad design with future implications.

I did notice that but I didn’t care either. This is toy bot after all, and my aim is not get a shipable product, but to teach him. I have put a lot on his plate already — from setting up the IDE, github.com account, doing commits, managing a python project with uv, etc.

The MVP

For next few days I could not keep up. My wife took over and acted as a product manager. She would ask my son for features, perform tests of the bot, and debating about what to do next.

She was also the moral keeper when things would not go as smoothly as in the beginning.

They managed to get the bot up and running at home, create travels, attach files (plane tickets, hotel reservations, etc.) to each of the travels.

The first screenshot of the MVP

The first obstacle

The command /startnotifications, as cooked by the GPT never worked. We wanted the Bot to send automated reminders about planned trips the day before they start, and send us back all the important files (tickets, bookings, etc.) so that we had them at hand.

My son could not make it work. He provided the GPT with the full traceback of errors, and the GPT would double down and give him back code that didn’t work. It insisted on get a context.application.job_queue which is clearly None:

Traceback (most recent call last):
  File ".venv/lib/python3.13/site-packages/telegram/ext/_application.py", line 1311, in process_update
    await coroutine
  File ".venv/lib/python3.13/site-packages/telegram/ext/_handlers/basehandler.py", line 158, in handle_update
    return await self.callback(update, context)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "main.py", line 64, in wrapped
    return await func(update, context, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".venv/lib/python3.13/site-packages/logfire/_internal/instrument.py", line 93, in wrapper
    result = await func(*func_args, **func_kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "main.py", line 182, in start_notifications
    current_jobs = job_queue.get_jobs_by_name(str(chat_id))
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'get_jobs_by_name'

Since at the time I didn’t have the time at the moment, we all decided we could fix that later, and the moved on to other features.

The second obstacle

The next obstacle was the hardest. The bot would not fail with an error, but it would not work as expected.

At this point the single file main.py had become a huge melange of vibe-coded ConversationHandler, MessageHandler, several intermingled state machines, and the familiar (and mostly useless) comments that all GPTs generate.

This was the breaking point. Where my son met its own limitations and we could no longer keep feeding the GPT with that 1000 lines of code and expect it to figure out the right approach for this stuck state machine.

There is the argument that the free GPT is not powerful enough and a more powerful one would be able to debug this. There is also the cost/benefit perspective. This is a non-profit project and my son cannot throw his allowance into an expensive tool he cannot use profitably.

The way out

I’m now taking a more active role in the project again.

ASK_NAME, ASK_DESTINATION, ASK_START_DATE, ASK_END_DATE, ASK_DOCUMENTS = range(5)
ASK_TRIP_ID_TO_EDIT, ASK_FIELD_TO_EDIT, ASK_NEW_VALUE = range(100, 103)
ASK_CONTINUE_EDIT = 103
ASK_AFTER_EDIT_OPTION = 104
INFO_NAME, INFO_LASTNAME, INFO_BIRTHDATE, INFO_CERTIFICATES = range(4)

Being me, however, I also started to refactor this and the first thing to do was to use an enum.StrEnum.

This didn’t actually solved the bug and therefore I’m taking a look at the Telegram’s API and concepts.

The learning

The important take-away for him is that he needs to use the AI as tool, but he also must gain more understanding about what the AI is doing. Taking myself a role model, I could spot some issues (like the repeated state identifiers) in a single glance, and he could not see that mistake.

There is still (and I think there will always be) value in learning the basic skills needed for programming.