Recap from last week
Agora has launched (the softest launch).
I’ve just completed final testing and plan to start messaging users this week.
I also wanted to see if I could build an iPhone app and use Cursor for building apps, so I’ve built one called Morning Pour which is an AI latte art coach.
AI was still making me mad at times (more on this coming).
What have I learned?
Every model I’ve used is getting better, but has an equivalent uptick in confidence.
This confidence aggravates me more than anything else, because it just goes ahead and makes code changes on any question you ask.
First thing learned - this anger is wasted tokens, even though it briefly feels good.
I was watching the thinking mode of the AI and it spends a lot of time (and tokens) trying to sort out your anger, rather than the problem.
I need to get hotter on using the ‘ask’ and ‘plan’ buttons that each model provides. This is the only way you can genuinely have a chat. Even prompting the AI to ‘not go ahead and make any changes’ will still not be able to suppress the confidence.
On that, it’s important to always read the thinking notes.
The way you can best understand where your tokens are going, and exactly what the model is thinking, is by reading what it’s thinking.
Every AI model makes this pretty easy, although I found Cursor didn’t make it easy (or even possible?) to read this. If anyone knows a way for Cursor to show me this, please let me know!
Do not rely on AI to write your .md files.
It’s very common to hear people tell you to just get the AI to write your agent files to educate the AI on how to operate, but I’ve found this to be terrible advice.
AI models are genuinely terrible at writing them, even when they have full context of the project. Just like they’re not good at writing to sound like a human, they can’t explain things without going too high level.
It means any agent-type bot you’re trying to set up loses track of what it’s supposed to do very quickly.
Next up
Going to be heavily focused on growing the product and getting feedback, so will be working on gerald.md my growth agent, as well as developing some way the LLMs might be able to help with growth strategy and company analysis (competitors and similar companies). I’ll share what I can next week.