Spent the weekend playing with claude code for my first nontrivial thing. It's pretty wild. It feels so much like working with a super green junor dev. A lot of the same coaching strategies seem to apply well. Also, I'd forgotten how it's kinda fun to teach good debugging practices! One major difference — I tend to hit the context window limit after about an hour, and have to start up a new session (because compaction sucks), which means I (seemingly) need to do a bunch of coaching to make sure claude is always externalizing its "thought" process to files so that the next agent can pick up where it left off. This sucks, or I just haven't figured out a good foolproof way to do it. The nontrivial thing — I'm making a new version of my depth camera, but using gaussian splatting. So I'm using this as an opportunity to: * learn claude code * learn gaussian splatting * learn more swift

I look forward to hearing more as the process unfolds 😊

There is a big difference between LLMs working on lowest common denominator code architectures and novel architectures and patterns. FWIW, in the coding agent I use, I seldom allow my agent to pass 30% context window.