Claude Code can just do things

I’ve had two 🤯 moments using AI coding tools. The first was when I tried Cursor and its tab-complete feature started reading my mind as I was coding. The second was when I fired up Claude Code, asked it to complete a task in my codebase, and watched it work through the problem in the same way that a developer would.

After my inaugural Claude Code session, I noticed that I had spent several dollars in a few minutes 😰. AI Agents are producing better results when they aren’t optimized for cost saving. My brain is not wired to optimize in the same way. Watching money being sucked into the void became a mental blocker for me.

Subscribing to the $100/mo Max plan allowed me to start experimenting with Claude Code without the worry of overspending. Since then, it has become a pillar of my local development workflow.

I think of Claude Code as a tool for doing. I use it to do all types of things:

• Resize and optimize images: It can use sips or ask to install ImageMagick with brew to perform various image operations.
• Load Android binaries onto a device: Provide the path of the aab or apk file and it will use adb to find a connected device and install the app package.
• Troubleshoot docker images: It can examine container filesystems and configurations to find production build issues.
• Source code diving: It can trace through source code and provide details in a more centralized and digestible format.

Recently, I spent half my day tweaking and rebuilding a mobile application. At the end of the day, I asked Claude Code to “Count the number of times I ran commands starting with ‘eas build’ today”. The answer was 14 😫. This useless little data point brought me joy, but I would never have spent the time to script a solution for it by hand.

If I know that something can be accomplished through the terminal, but I’m not sure what the command(s) are, I ask Claude Code. If I’m trying to accomplish something advanced, I’ll have it read the man pages first.

Much like coding of days past, working with AI generated code is a daily practice for me. It’s a skill, and like any skill, you will perform it better with lots of practice. Part of this skill is recognizing areas where the AI falls short and what workflows are effective relative to your use case and constraints.

When working on software that isn’t throwaway, I keep it on a tight leash. When I ask it to write code, I’ll try to make sure that it has all the required context, provide patterns and conventions for it to follow, and give specific instructions about expectations and desired outcomes. Built with borrowed hands dives deeper into these strategies and highlights some of the common challenges.

It can also help with brainstorming and working through writer’s block. Getting some code down as a starting point can be a great way to get the ball rolling. Be careful not to fall into the trap of letting it think for you. These tools are not a substitute for human thinking and creativity.

If you haven’t tried Claude Code, give it a week with pay-per-usage or the $20/mo plan. Read through the Common Workflows and Best practices for agentic coding for tips and best practices.

Don’t limit yourself to basic code generation tasks. Delegate as much of your workflow as possible. Claude Code is a doer. What it can accomplish is up to you to discover. You won’t get an extra four hours by the pool every day, as TV commercials suggest, but with practice you will get a hell of a lot more done in a few hours than you used to. Experiment and have fun along the way.

---

Additional resources

• I Shipped a macOS App Built Entirely by Claude Code

Random things I’ve used Claude Code to help with

• Create a launch agent to make sure my ssh key is loaded when my laptop starts.
• Cleanup login items and other leftover artifacts on my machine from deleted applications.
• Remove a blank page from a PDF.