After a couple of weeks using Codex, I am pretty happy with it as a step-up improvement over Claude Code. It tends to do the right thing more frequently and without interaction, and the sandbox model prevents continuous interaction over permissions. Still not a perfect tool, of course, but I am pleased. #AI

Enjoying how trivial it is to improve all my development tooling and environment by having AI just write the tool I want and haven't made time for. Latest example: macOS xbar plugin that tracks both my open PRs and also review requests. It uses magick to generate nice icons that sit in the tray like this (the two numbers on the left):

Our free credits with Claude (Anthropic) ran out and so I've been using Codex (OpenAI) for the last few days. It's much mellower than Claude in expression, but seems maybe more competent even. So far I am impressed.

I'll be writing a blog post about this, but at Mozi we just open-sourced a tool I wrote that I've been using a lot: https://github.com/mozi-app/mermAId . It's a Mermaid live editor / MCP server that is made to interact with LLM agents while planning. The agent can push and pull Mermaid diagrams from the editor. The editor has vim bindings support if you want to use them. On macOS it's a native app, on Linux and Windows it runs a server and opens the UI in the browser. There is an accompanying CLI script for those who prefer that over MCP. The details are all in the README.

I spoke at the Augmented Software Engineering meetup two weeks back about 6 months of agent-assisted coding with Claude. #AI #Agentic #ClaudeCode

Slides are here

Video is here

The thing with the AI tools is... they are tools. It's not a sentient thing. It's a tool. And you have to learn how to use it if you want to get anything out of it. The fact that it's being sold as not-a-tool is the main issue with regard to utility. "Your virtual co-worker!" "Anyone can use it" "Just use natural language", etc. Nonsense.

I just asked Claude Code to diagnose an issue in code it wrote. I told it to also use Gemini on the CLI to do the same analysis. Claude was to discuss the issue with Gemini until they had a consensus on the problem. This actually worked! Kind of nutty, but the outcome was a good plan to fix it. Better than what I was getting from Claude alone.