Claude Code's latest & most important feature
And why it's not Sonnet 4.5 for me...
My inner control freak hates using AI agents to code.
Last week, I built a learning app to accompany a course I’m developing.
Since this was an experiment, I let the coding agent run free (what some call ‘YOLO mode’). By the end of the week, I was lost. The app ran and seemed to work, but I’d never launch it. I couldn’t make sense of my own code.
Today? Anthropic added a critical feature: checkpoints in Claude Code that could help with this (which is more important to me than the new Sonnet 4.5 model that everybody is talking about).
Anyway, I rebuilt that app from scratch. Here’s what I think are the 3 things in my coding environment that placate my inner control freak.
1️⃣ Clarity - Track changes in diff
This was what made Cursor THE choice for me a while back. They had the best diff view (i.e. tracked changes in code). I prefer to step through and accept changes one by one in the diff view. No vibes involved.
Most Gen AI coding providers have caught up. The baseline has risen here, which is great for everyone. The key idea here is that if you can’t see what’s changing, you’re not coding. You’re just jabbing at your keyboard.
2️⃣ Control - Create checkpoints
This is what you need when you go YOLO when coding. Letting the agent make change after change with no way back is not just a bad idea. It’s a really dumb move.
To address this, we could always be disciplined with git commits. But let’s be honest. It’s a pain when using Gen AI to code. I’ll either do too little or too much.
Cursor has had checkpoints for a while. GitHub Copilot also added this. And today, Claude Code launched their checkpoints feature with Sonnet 4.5.
Checkpoints give you a safety net. Experiment boldly. Roll back instantly. Simple but life changing (at least to me).
3️⃣ Critique - Review the changes
There are specialized tools now. CodeRabbit does AI code review. But you don’t need another tool to start.
In GitHub Copilot, just select the code and choose “Review.” Get a second opinion from the AI itself. Challenge its decisions. Ask ‘what could go wrong’. It’s fun. Trust me.
My evolving coding setup.
I started with GitHub Copilot when there weren’t many alternatives. Moved heavily to Cursor during its heyday.
Now? I use:
Claude Code for initial experimentation and planning, and proof of concept blitzes
GitHub Copilot for understanding, refinement, reviews.
So 3 “Cs”. Clarity so you understand what’s changing. Control so you can experiment without fear. Critique so you catch what’s missing or wrong.
#AIinWork #GenerativeAI #CodingWithAI #ClaudeCode #Sonnet45