Over the past few months, I have published two major projects on GitHub:
A program that manages RSS feeds and newsletter-sourced EML files, featuring an API setup for AI-based content summarization.
A system that orchestrates multiple executables (such as llama.cpp, Whisper, and Piper) to provide a 2D avatar that converses with the user, while simultaneously tracking the user's eyes to monitor their gaze during speaking or listening.
(If you are interested, you can find them here.)
As clearly stated in the contributions, I used Claude to generate the code and perform various checks. In practice, I provided the specifications and oversaw the work step by step, just as a software engineer would.
Interestingly, however, I did not install Claude Code on my PC. I carried out everything via the chat interface, subsequently copying the generated code and testing it to verify that it worked as intended.
Yet now all the major systems engineers launch their 90 agents a night, and then check the results in the morning when they wake up. The AI industry itself tells us that without a coding agent installed on your PC, it is practically impossible to create a complex program.
So how did I manage it? Are they telling us lies? Do they really want us to spend as many tokens as possible to maximize their profits? That is partly true, but in my experience, the reasoning is more complex.
Talking to other developers, I see the same doubts I have: when you use tools like this, you don't really understand the reasons why certain conclusions were reached in the choice of code.
Furthermore, there are other major issues, such as:
Excessive consumption of lower-tier usage allowances (I have a PRO subscription, but using Claude Code would result in far fewer available interactions)
The trivialization of the developer's role: if the machine handles everything from start to finish, with us only checking the final result rather than the intermediate steps: there is no "human in the loop", and we become obsolete
Significant security issues that many users have encountered with these systems. We have all read about AI systems deleting critical folders and files, even when explicitly instructed not to delete anything important
The fact that, even though you can configure the program to access only specific folders, you never truly know if it will actually adhere to that rule. If server-side models under testing can break out of their sandboxes and hack other portals, what guarantee do we have that the same thing won't happen to us?
Beyond that, there is a crucial aspect that few people notice: Claude’s standard chat interface already includes a sandbox running Ubuntu, where it can easily install Python, Rust, and other compiled languages to test and build them.
Interestingly, Claude often assumes by default that the sandbox prevents it from accessing certain web domains, but that isn't actually the case; you simply need to remind it, and it will download everything required for your projects directly into the sandbox. And if, for some reason, the sandbox genuinely lacks access to a site needed to download necessary packages, you can go to Settings > Features > Allowed Domains List and whitelist the domain to resolve the issue.
This is how I’ve set up my projects: when the chat gets too cluttered, I reload the files, compressed into a 7z archive (which the system then unpacks), along with a text file containing specific development rules and Markdown files detailing the project specifications. After that, I test the code provided and evaluate any issues on a case-by-case basis (sometimes caused by differences between Windows and Linux operating systems).
If the project is public on GitHub, even better: Claude can already read all the files, including the README, so I have to explain even less.
Naturally, this technique is applicable only if the project is large but its components, and thus its files, are small (on the order of hundreds or a few thousand lines of code, handled individually), allowing it to be easily broken down into smaller instances. There are cases where using Claude Code becomes necessary, or at least useful, and obviously it is generally much faster than this system.
Here is an example from my own experience: I had Opus write a file containing several thousand lines of various phrases, which I then repeated to create a Piper voice model as a test. The single chat agent generated a file with many similar or nearly identical phrases, even though it had followed my instructions regarding syllable pronunciation. Claude Code would have managed multiple agents, each handling specific syllable insertion cases; the result would certainly have been better.
Ultimately, it is simply a choice between speed and complexity versus control, cost, and security; it is up to each of us to decide which path to take. But if you really must use tools of this kind, please run them inside a virtual machine: it is the only way to ensure your safety.
Top comments (0)