Teaching Claude
I had the horror realization that I need to start interviewing soon, and I remembered that my memory sucks. SOMEONE is going to ask me my thoughts on AI, and I already knew my answer: It's super useful for increasing my speed at things, but I need to watch its output closely, because it does some junior dev things I don't like. The answer I wouldn't have off the top of mind is "give me some examples" - so that is what this page is for. Most of it isn't earth shattering, just little isms.
I use PULL_REQUEST_TEMPLATEs heavily, customized per project/repo/task. I generally have a Summary section, and even with a comment to the effect of "this does not need to be fully fleshed out - this is just a quick glance", Claude, when using the template, would write BOOKS about what was changed.
I updated my comment to be "Be concise - Files Changed is source of truth" seemed to help somewhat.
This was discovered while having Claude knock out a post-checkout git hook for me which would ensure that worktrees use relative paths for gitdir. I've started doing most of my dev work in a VM which has Claude, with a mounted directory to the repo I am working with. I absolutely do not yet trust it with write git access, so I've generated some readonly PAT for the VM to make it useful. When creating a git worktree, gitdirs are all absolute paths, which is unhelpful on a mounted dir in a VM. (This will eventually be added to my template-repo).
