1.1 Introduction. Who is this guide for, and why is it focused on private and corporate development with LLMs?

Hi, everyone! I created a new blog and decided to start with the most relevant topic today: product development with AI. This guide is intended for a broad audience: everyday developers, people just starting out in programming, and those who have spent the last few years asleep and paid little attention to AI. I’ll try to explain all the important aspects briefly and clearly, and share recipes and tools for writing maintainable, extensible, production-ready code that people can understand.

Why will we be focusing on corporate coding standards?

When you’re vibe coding and building your own application, you’re limited only by the boundaries you set for yourself. And all the problems that arise as you develop your project are yours alone to deal with.

If we’re talking about a company with its own product that is constantly evolving, and whose team consists of at least several developers, the first problems will arise very quickly.

If we view a company as a well-established profit-generating system, the development department will be a subsystem needed to ensure the operation and growth of all the company’s websites.

For this subsystem to function reliably, the following requirements must be met:

• Developers will be hired and let go, so this resource must be easy to replace without any critical consequences for the product. It must not be the case that only the most senior developer understands all the hacks in the code—the very hacks they introduced themselves.

• The application should have an architecture and conventions for writing code and conducting reviews, testing procedures, and a long list of requirements that must not be violated. Ideally, everything should be documented so that new employees can get up to speed faster and more easily.

• A well-established development process and workflow are essential. Even if a task is completed by changing a single line of code, a colleague will conduct a code review, and a tester will verify that everything works as expected. This process is the same for both small and large tasks. All changes are recorded in the task tracker. This makes the result more predictable and stable, and also helps us investigate later if something breaks or goes wrong.

As you have probably already realized, we will be building a system that makes it possible to achieve predictable results and work on the product long-term without a massive headache, while bringing the entire development-review-testing chain together under one person who will work as part of the team and strengthen it. This remains just as relevant for solo developers.

Problems with subscription-based cloud models

We’ve been quite literally hooked on cloud-based models—they’re convenient and, for now, relatively inexpensive.

The biggest problem is your dependence on third-party infrastructure. A single political decision, and you could be deprived of a tool you’ve grown accustomed to and can no longer imagine working without. Under the pretext of sanctions or for any other reason.

The second problem is unreliable performance and a black box under the hood. If you work extensively with cloud-based LLMs, such as Claude Code or Codex, you’ve definitely noticed huge periodic drops in generation speed during peak hours. But there’s something even worse: you may have Claude Opus enabled, while in reality an extremely weak model is doing the work in low-effort mode, and tasks that used to be completed in one or two prompts now don’t get completed at all. At times like these, you can’t perform your work tasks the way you used to or at the same speed. And there’s no one you can hold accountable. Pay up and don’t bitch. Reddit users also frequently point out that when a new model is released, the older models become noticeably dumber. There’s a good article on the subject— https://habr.com/ru/articles/1023020/

The third problem — this is privacy. If you are one of those people who "have nothing to hide," then congratulations: any state, as a supersystem where laws are churned out every year that allow it to dig ever deeper into your underwear and use the information already collected to take your money, will put you first in its sights. AI will accelerate this process considerably. When it comes to commercial companies, the risks increase significantly. Wherever there is a lot of money, there is a lot of attention from various participants in different organizations, as well as from others who want to seize those resources. And if you believe that the data sent to Anthropic and OpenAI is inaccessible to anyone else, you are very naive. This is extremely valuable information, provided voluntarily.

Advantages of local open models, using Qwen 3.6 27B as an example

The biggest advantage is stable performance in the morning, during the day, and at night. The model does not become stupid in the evening; its speed is always the same.

Of course, local models are not as smart as giant cloud-based ones. But there is also a less obvious advantage here. You plan the task more carefully, breaking it down into smaller ones that even a very stupid model can handle. You will not be able to vibe-code recklessly; you will have to interact with it like an engineer, knowing its weak points. As a result, you will write code more deliberately.

I claim that the models Qwen 3.6 27B is quite enough for high-quality product development. It’s a dense model, with all 27B parameters active simultaneously. And it can be run locally on a graphics card with 32 GB of VRAM or on a Mac with Unified Memory.

I ran an experiment. I had a task for a project to implement a dark theme. I first used Qwen 3.6 27B, and then GPT 5.6 Sol. In other words, I completed the same task twice using different models. Each implementation took me about two hours. The results were similar. But I preferred working with Qwen—it wrote code faster and didn’t get stuck intermittently. The result was a pull request with 110 modified files. Doing this task manually, without AI, would have taken about five working days.

The boost is out of this world! Let’s move on!