Claude Is the Answer

I have been keeping an eye on agentic coding tools as they slowly emerged, and I have continuously tried them out. Previous iterations have been fun to use, but nothing I could see myself using as a senior engineer.

Until now!

If you don’t bother reading on, my message to you is simple: You must try out Claude Code with Opus 4.5, and strive to integrate it into your development workflow for production-ready applications. This is a message from a highly critical and quality-focused senior software engineer, who has previously been unimpressed with earlier iterations of agentic coding.

Background

Recently, I revisited OpenAI’s Codex (GPT-5.1) through our Plus subscription and, for the first time, it actually produced some code worthy of releasing. This naturally sparked my interest in revisiting Claude Code, which is commonly mentioned as the best tool for agentic coding.

I was highly impressed with both tools’ ability to produce code that was actually correct. However, any senior engineer knows the importance of (human) maintainability and simplicity in code. These aspects are where the two tools seem miles apart, and using the correct tool can, in my experience, make or break your desire to use these tools on critical codebases.

Codex

To put it directly, Codex often produces correct code (read: code that makes your tests go green) but completely lacks the ability to produce simple and (humanly) intuitive abstractions. Despite DRY principles often being wrong, Codex seems to avoid them at any cost. It almost deliberately avoids creating reusable abstractions, and even when directly instructed to do so, it fails badly in my experience.

In summary, Codex feels like the traditional consultant who focuses too much on fulfilling requirements and ignores any measure that might slow down the process of reaching that end goal.

Claude

As Disney’s Aladdin put it: “A whole new world…”. I simply cannot emphasize enough how impressed I have been with Claude Code’s ability (Opus 4.5) to produce production-ready code over the last week. Using Claude Code feels like hiring a brilliant junior engineer for $100/month who produces code at five times the speed of a typical human engineer.

Your primary task becomes instructing it on the changes you want and providing guidance on the abstractions you find ideal for the solution it should produce. In contrast to Codex, Claude Code does (i) actually produce code you can—and want to—maintain, and (ii) has the ability to simplify code by creating useful and reusable abstractions.

For example, this week I produced a high-quality internal package that would have taken approximately one month of manual development time, within four or five working days using Claude Code. I followed each change closely and was constantly impressed with the proposed code changes produced by Claude throughout the process.

Where does this leave us?

In the past, I have maintained what I would deem a healthy skepticism toward agentic coding tools. However, I must admit my recent experiences with Claude Code have left me partially speechless, and I am now fully on the agentic coding bandwagon. I have not had this much fun coding since I discovered Ruby on Rails in 2009.

At the expense of sounding arrogant, I have for the first time experienced agentic coding tools producing somewhat complex code that I actually liked on a consistent basis. If you have not tried these tools yet, you HAVE to try Claude Code.

Moving forward, these tools place an even greater emphasis on the ability to read and understand programs. You will find yourself spending most of your time reviewing code changes, with little time spent iteratively examining package documentation and running tests. For senior developers like myself, these abilities come naturally. However, for future junior developers, this presents a challenge, and I believe we need to revisit how we approach computer science education.

This is the “calculator moment” for software engineering. Just as computing math formulas became trivial with the calculator, producing code from human-specified requirements will largely become trivial as well.