jesseduffield
What is your AI agent constantly doing wrong?
My colleague and I have found AI is consistently falling short on the following things. I'd be interested to know whether this is a prompting issue or a model issue (we use claude 4.7):- comments being too verbose, and making reference to the previous code. Also just using phrases that no human would ever use (e.g. saying 'gate', 'belt-and-braces' a lot)- over-using default arguments in method...
On the other side of the spectrum is more wholesome long-horizon activities like a challenging side project, career progression, or fitness goals. There's certainly an element of variable ratio reinforcement in all of these, but because the rewards are so much more tangible, and you get to exercise more of your agency, these activities generally feel quite meaningful on reflection.
Playing pinball is somewhere in the middle, probably on the cheaper side of the spectrum. Introspective people can generally reflect on a session and decide whether it was a good use of their time or not.
I really think that 'how do you feel after a long session of this' is a good measuring stick. Very few people will tell you that they feel good after a long session of social media scrolling or short-form content.
Another good measuring stick is 'do you want to want to be doing this?'. I want to want to go to the gym and gain 10kg of muscle. I do not want to want to spend hours on tiktok every day.
Always Bet on Text
AI Tinkerbell
ChatGPT 6 Is a Jerk
AI is stupid, just like me
The Future of Redwood Launches Today
1M Lazygit downloads, one embarrassment at a time
I want my AI to get mad
Add More Kafkas
If you're out of ideas, build Memeware
Don't Let Go (of compile-time knowledge)
Lazygit creator interviews DHH (creator of Rails)
Stop selling me trichotomies!
Life Without the Front Page
Lazygit creator interviews 'Uncle' Bob Martin
If any of the above questions have been or your mind too, you might find this interview valuable.