Like many other developers, I recently started using some LLM-based AI systems as helpers for coding in a few languages. I'm not a fan of VSCode, and I prefer a more traditional approach to coding: I hate to cope with code completion servers and use one of my preferred editors, Vim or Emacs. Navigating by tags is more than enough for me. That said, this is the summary of my current experience in the new world of AI-aided approach to coding (i.e., AI-aided development or AIAD for brevity).
There's a clear divide among developers when it comes to AI tools: some love
them, while others are more skeptical. If you're keen to delve deeper into this
topic, Salvatore Sanfilippo, also known as
Antirez, shares some insightful perspectives on his YouTube
channel. He provides comprehensive
evaluations of the leading AI models and systems, offering a balanced view of
their strengths and limitations.
Based on my experience, while both Anthropic and Google top models are equally good enough for use, a big difference is due to their relative UX, which is incomparable among various systems. Luckily, things are continuously changing in those areas. Copying and pasting without a clean and accessible per chat/project assets catalog is definitively a pain. The most effective approach is sharing pieces of a git-based code base, which is not always possible with all AI chatbots. Even Salvatore shared a few impressive results about his experiences, and I can definitively assert that, at least currently, some of those systems can be quite helpful for some tasks if (and only if) used with a grain of salt.
First and foremost, the most effective use of AI tools starts from very circumscribed tasks to perform in a step-by-step approach after a well-stated creation of the context to make choices. In other words, any model must be conducted by hand in the right direction. It is crucial to use what Claude AI defines as projects to create a clear context with a general detailed architecture description and key assets for orienteering the linguistic model. It is also essential to maintain order meticulously. Being verbose enough and precise is the key point, and that's typically what a senior profile should be able to do. Each resulting asset needs accurate reviewing and testing; it is simply impossible to assume that a simple change in human terms corresponds to perfectly valid changes in AI assets. What is easy for us could not be for the AI model and viceversa.
The review should be both stylistic and functional because complicated programming patterns could be completely unmatched by the AI model. That is specifically true when documentation about APIs is unprecise or missing the point. The models even tend to generate redundant code or hallucinated code snippets that cannot be compiled or interpreted. Sometimes, the AI model generates incredibly good code at large but presents silly or subtle oversights. At the end of the day, there is not always a gain in development time; it is only something different: you spend more time reviewing and fixing bugs than writing code from scratch. It is essential to cope with the limits of current AI systems to develop a handy way to compose multiple parts together and keep track of improvements and fixes in multiple iterations of the process. During iterations, it could reach the limits of the AI context and get incomplete or truncated files, and it is fundamental to have a way to regain the path with minimal effort. In other words, one has to design a well-formed idea of the resulting products and all intermediate artifacts. That's the reason why seniority and experience are fundamental to using such tools effectively: sorry, Mr. CEO, engineers are still here to stay and being paid for good work.
Tasks that can be easily covered by current AI models are:
- Creation of general documentation and/or comments in existing code.
- Creation of simple tests from a structured codebase.
- Translation of code from one programming language to another, including documental formats, such as XML, JSON, and YAML.
- Creation or improvement of auxiliary tools and boilerplates.
- Reviewing and improving existing code by steps.
- Help in identifying common issues and possibly interesting features.
The average results are good enough for an initial base of successive analysis and development, and I found the AIAD especially helpful with activities that I judge seriously dull and time-consuming. It's a relief to have AI tools that can handle the whole web front-end-related activities, which I personally find less interesting.
Using multiple AI models to cross-review a code base and possibly prepare multiple improvement alternatives while purging the obvious misunderstandings is also an interesting opportunity. At the end of the day, I judge AIAD as effective enough, but for the most significant limitation of having a not-too-extended context size (at least with basic pro/premium profiles). This intrinsic limit is the source of major problems in the UX, along with still rough or incomplete interfaces: if you need to review and modify a large code base, you can easily crash against such limitations and have to apply intensively a divide-and-conquer strategy to govern hundreds of thousands of LOCs (but diving into any unknown big code base is anyway so different?).
Of course, the correct approach is understanding that its use changes programmers from pure creators to reviewers. In those regards, I generally consider AIAD a Stackoverflow with steroids: anyone who used SO in the past found valid and interesting answers to questions along with perfectly misleading suggestions, and that's not different, and mostly better.
Disclaimer: I mostly used Claude and Gemini Pro, much less ChatGPT and Deepseek due to their intrinsic limits for UX.