▪ claudecodedaily

Windows users shouldn't need a PhD in VM setup just to run AI coding agents. Building: browser-based terminal → VM → Claude Code ready to go. Researchers: semantic search MCP for academic papers & skills for lit reviews added Friday night, debugging the API proxy layer so

IDEs are becoming app runtimes After spending most of this week building a few Claude Code skills, I don’t think writing the agent harness for most tasks is a good use of time. I think the harness made sense before the latest models but not anymore - give me Claude code every time. With Claude Code, your agent loop is also your IDE, and developing “within” the same loop your app uses creates a quick feedback cycle that can unblock most AI app development hurdles. You can quickly see its behavior (over many tool calls), steer it, and it can update prompts within your app, help you label some examples, build and run an eval, you can have it write tools for itself or update its skill md file. The fact that you can do the coding, prompting, evals/human feedback - all while feeling like you’re conversing with a peer - it’s just hard to see how this doesn’t become the main way to develop AI apps. I felt like I built intuitions for when you can bake in robustness within your code or keep the squishy human workflows to the markdown file - and then rely on a very capable model to fill in the gaps - I think there is a lot there that is still being figured out… and I guess it’s trending towards being more squishy. It wasn’t until I sat down and tried to automate parts of my job did I realise how squishy they were. Even just triaging feature requests - I have to read the entire context, maybe watch the call, maybe talk to the customer, talk to an engineer - I realised that I could speed it up but I had to codify (in markdown) where I wanted help versus where I knew I couldn’t get it. While there’s still non-squishy things that are better suited to code, I’m still not sure how to pair engineers with the domain experts so that you can both sit “within” the agent loop together, because to me, that feedback loop is the key to moving fast while these models keep getting better and we’re all building on quicksand.