I recently tried using Minimax (M2.5) to build a few features. I found that for tasks that aren’t too complex, it can do the job very well, as long as you describe the task clearly and let it test after implementation.

For example, I asked it to add an “insert image by drag-and-drop or paste” feature to my Markdown tool Mai. After some analysis, it proposed a plan and implemented it. It didn’t handle the screenshot input correctly, but made it work after a few iterations.

However, I didn’t fully trust it (maybe due to I was not familiar with it yet), so I asked Claude Code to review the implementation. Claude was straightforward and gave me five or six criticisms. They seemed solid. Most of them weren’t about the feature itself, but about the overall app architecture and macOS best practices.

So I asked Claude Code to fix them. Here’s the interesting part: Claude code broke the screenshot paste function that used to work. After a few more iterations, Claude fixed that of course.

This whole experience clarified something for me: for straightforward features, plan execution or building isn’t the main differentiator anymore. Claude Code or Codex might not be dramatically ahead for small to medium tasks if they are precisely described. Other models such as minimax can often get to a working version too.

What feels truly different is planning.

Claude Code and Codex seem much better at placing a request in a larger context, which changes the trade-offs they pick, and the plan they propose. This planning power is especially crucial for complex tasks, where vibe planning is where the real productivity gain is.

But even with Claude code or codex, human out of the loop is still not there based on my practices. You still need a human in the loop to control complexity and handle details efficiently.

Every developer has limits. None of us can be an expert in everything, so we naturally frame problems and search for solutions inside the space we already know.

AI has seen far more patterns and possibilities. It can explore a much larger solution space and often comes back with better options.

Our job is to ask for directions and suggestions, learn quickly during the planning stage (with AI’s help), and for now keep the final decisions in human hands.