How I actually work with AI: my tools, methods, and a few honest observations
Some time ago, sparked by the OpenClaw boom, I started closely following interviews with Peter Steinberger. He spoke several times about what his daily work with AI looks like: what tools he uses, what methods, what his workflow is. I found it an interesting format and decided to describe the same from my own perspective. Maybe it will inspire someone, maybe at least one tool will find its way into someone's daily work.
If you're curious about how I actually turn these tools into results and deliver projects, feel free to reach out. I conduct training and implement complete AI workflows in software development organizations.
What I use daily
For programming with AI, I primarily use two tools: Claude Code and Codex. I'm a big fan of Anthropic's models, and Claude Code is currently, in my opinion, the most advanced tool of its kind. However, OpenAI is clearly catching up. Both their models and Codex itself are becoming increasingly mature and useful in daily work.
I previously used Cursor and Windsurf, but six months ago I permanently switched to programming exclusively from the terminal and have no intention of going back.
When someone asks me which tool is the best today, I answer with humor: I don't know, because I haven't programmed today yet. The pace of development of these technologies is so enormous that to truly stay current, you'd need to use everything simultaneously, which is obviously impossible. I try to at least follow the leading ones: Claude Code, Codex, and Gemini. I'll write about the last one separately, as it plays a specific role for me.
How much I consume
I program with AI daily, for several and often over a dozen hours. Sometimes, when I need to deliver a project demo and execution quality isn't a priority, models work autonomously around the clock until they complete the task.
For this reason, I maintain two Anthropic subscriptions (Max 20x) and two from OpenAI, and even then there are situations when I exhaust my limits.
If you're curious about how many tokens you actually consume, I recommend the ccusage tool. It generates very detailed usage statistics and it's really worth seeing. My record-breaking February 2026, exclusively from my local machine and exclusively through Claude Code, amounted to tokens worth $5,000 in total. On top of that, there's what I consume by installing tools directly on servers, but I haven't counted that.
Methodology: BMad
Over the years, I've tested various work organization methods. I started with simple Markdown files, tried OODA, GitHub Spec Kit, and several other approaches. Ultimately, I'm the biggest fan of BMad, and it's exactly this method that I've implemented and used to train developers in several organizations.
BMad provides tools for every stage of the software development process. It starts from the idea: research, brainstorming, architecture design, brief creation, UX planning. When it comes to implementation, the method guides through the entire process, from sprint planning through building epics and stories, all the way to Code Review and retrospective.
Several things make me particularly fond of it. These are methods well-known to developers, so they don't introduce additional confusion. They bring structure and order to the entire process. And above all, AI has precise instructions thanks to them: what stage the implementation is at, what needs to be done now, what the next step is.
Important note: BMad is not another tool to install. It's a methodology based on agents and skills, and deployment is practically instantaneous.
If this makes sense to you, feel free to reach out. I'd be happy to train your team and ensure the entire organization successfully implements this flow.
I write less and less, I speak more and more
One of the directions in which my work is evolving is replacing writing with speaking. I started extensively using MacWhisper, and it truly changes the way you work. You can quickly dictate large amounts of text without using your hands at all, you can test something and speak at the same time, and the whole process of taking notes becomes much simpler because you don't have to switch between windows.
However, recognizing certain limitations of this approach, I decided to create my own software for orchestrating the entire process. That's how Nupi was born.
Nupi
This is the first place where I publicly talk about this project. Apart from my friends, nobody knew about it until now.
Three years ago, talking with a friend, I described a vision of the future where the way humans communicate with computers would change. I said the time would come when instead of typing on a keyboard, we would simply talk to machines. That a programmer would be able to have a dialogue with a software development tool, plan with it, give commands, build applications with voice. I didn't expect then that I would contribute my own brick to that future.
Nupi is a tool that allows you to control any CLI process by communicating with it through voice, with AI as a partner coordinating the entire development process. It's free and open source.
What can it specifically do? It has memory and knowledge about your projects, so you can entrust it with tasks knowing they will be completed. It has built-in cyclicality, so you can ask it to report project status every half hour and it will do so. It's independent of the CLI tool: it doesn't matter whether you use Claude Code, Codex, Gemini, or another tool. If you say that main features should be built by Claude Code and Code Review should be done by Codex, from that moment on, that's exactly how it will work. Nupi will handle communication between these tools.
And it lets you work anywhere. You don't have to sit at your desk. You can talk about your code on your way to work, during a walk, or while driving. I've already found myself developing Nupi using Nupi while driving. A fantastic experience.
I'm convinced this is what the future of programming will look like. For me, it's already the present.
At the computer: lots of terminals
When I work at my desk, I operate exclusively from the terminal. Most often I have many windows open simultaneously, typically in a 3x3 grid, but it happens that I work in parallel on twelve or even more terminals at once.
When I focus on one project, I use git worktree. Most often, however, I work on several projects simultaneously, and it's really not easy. It requires constant focus and continuous coordination of many agents' work across many independent features. It can be exhausting. In my training sessions, I devote considerable attention to this, because I believe that underestimating this difficulty is one of the main reasons people become discouraged from working with AI.
Gemini and Repomix: when I need to see the big picture
Despite promising benchmarks, Google's models in my practice still clearly fall behind when it comes to programming. However, there is one area where they are irreplaceable for me: the one-million-token context.
Tools like Claude Code are optimized for quickly and efficiently finding relevant code fragments. These methods are truly advanced, but still fallible. Sometimes we need to look at the entire application from a bird's eye view: see the full architecture, understand how elements connect to each other. No tool will do this well if it doesn't analyze the project's entire codebase.
And that's exactly what I use Gemini for, combined with Repomix. It's a simple tool that saves all application code into a single file. It accepts various arguments to exclude selected fragments, respects .gitignore, and thanks to this, even with larger projects, we can often fit within the one-million-token limit.
The resulting file is very easy to attach to Gemini. The entire codebase becomes part of the prompt. We can ask for a Code Review of the entire application, architecture analysis, finding repetitive code for refactoring. AI as an auditor of the entire application delivers truly valuable insights, and I recommend this flow to anyone who hasn't tried it yet.
Finally: hardware
I work on a single MacBook that is slowly feeling the lack of RAM, and a replacement is probably coming soon. But I wouldn't trade my 49-inch ultrawide monitor for anything. In the era of working with multiple terminals simultaneously, it's absolutely essential hardware. It fits a truly large number of windows at once, and when you use it, it's hard to imagine going back to anything smaller.
If you're interested in AI-assisted development training or implementing BMad in your organization, feel free to reach out.