AI

I always enjoy chatting about how folks use AI and what tools they use. Partly because everything is changing all the time and you never know if you are missing out on the best tool ever, because let’s be honest, it’s easy to miss, given the torrent of AI-related information coming out every day. Opus 5.5 The GOAT, enough said! Chief of Staff / Firstmate Most of my work now happens through a firstmate-type session that manages other ‘worker’ sessions. The firstmate’s job is to talk to me and manage the other agents for me. These are not subagents but full Claude Code/Codex sessions. This really allows me to scale and do more. While I do lose some visibility into the actual details in the process, I can always jump into an individual session and chat directly if I want to. Which I sometimes still do, but not that often. ...

I spoke at AI Engineer Melbourne about “The Software Engineer Who Don’t Code”. It was a bit daunting to get on stage 😅 but it was also super energising to meet other smart and kind people who are thinking about the same problem space! I’ve always been a fan of the AI Engineer series and have watched many of the talks on YouTube. But flying to SF just for a conference is a bit expensive, so I’ve never really been to one (maybe it’s time!). When they came to Melbourne I knew I had to go. And speakers get in free! 😉 ...

I’ve been playing with 4.5 and Claude Code 2 for a few days now. And I’ve used it via Claude Code and via API in Zendesk App Builder. In general it seems better and faster and tool calls definitely seems faster compared to the pervious version. Also I’ve noticed Claude Code taking more notes. And I’m yet to notice any context anxiety Overalls it feels like an definite improvement. However, I struggle to quantify these improvements as a user. Like it’s better but it feels evolutionary not revolutionary. And it seems that Anthropic is listening to the community and watching whats happening among builders and then taking these lessons to the model training stages. For example, externalizing context/memory to the file system. This is great but I’m not feeling the AGI here, and it’s not a bad thing! :) ...

Low-fi wireframes are the best (these are from codex-cli)! They are: LLM and terminal friendly Low on tokens Quick & easy validation of UI before you move to more high fidelity mockups or just straight to implementation. Don’t think I’ve seen these much in Claude Code, but I’m sure Claude will oblige if I ask :)