LLMs below Opus produce a surprisingly high rate of malformed bash commands that have to be corrected iteratively or just shown the right way from the beginning using aforementioned multishot prompting.
For claude web app the best way to have the agent remember things is to have it open up a running todo list in an artifact that you ask for it to frequently update.
"This is where I would keep track of things that I always want included in any bash or other routine I write that is executed on the command line. I will add to it as I think of new items:
When listing database items in a list across the terminal screen, always take into account both the size of the screen as well as the line elements using a width that can consistenly show (if possible) all of the data across the screen (but truncate if necessary
All scripts should in the code show the version number as well as time and date of the latest version."