The Future of Small Utility Programs
Stop searching for the perfect CLI tool. Learn how to generate custom utility applications with AI in minutes. Real examples of 0-shot tool generation included.
Recently, I have been experimenting with the potential of using AI to create small utility programs and scripts. By small utility apps, I mean any programs or shell scripts that are up to 1000 lines of code. Programs like JSON manipulators, file converts, CI/CD scripts… etc. I have successfully generated a few CLI programs at an astonishing speed. Some were 0-shot.
- A VCF to Markdown converter
- A Jira issues exporter
- Invoices retrieval bot
Some of these cases might already have good CLI application. I have done some research and found good ones that, more or less, does what I was looking for. However, “more or less” is not exactly what I am looking for. It is usually about adapting a tool, even if it means compromising on what “exactly” I am looking for.
With the advancement of LLMs, I have realized that I no longer need to adapt. It is the programs that need to adapt to my requirements. This is a huge game changer. What I have concluded is that it no longer matters to try and find small utilities because you can just generate them.
LLMs has probably already seen all the programs and scripts in GitHub and more. If there are good ones, it can try to assemble one from existing programs by chaining them in a bash scripts. It can use their SDKs to build a concise Python or Go program. It can copy their code to build a tailored version. It does not matter how the LLM is going to generate the program. What matters is that it works.
When requirements for the program change, the utility app could be mutated by an LLM (probably a newer and smarter one). The utility app could be re-written again from scratch. I repeat, it doesn’t matter. What matters is that it works.
While this sounds easy, it is actually not. This particular workflow requires letting go of control. One should not care which programming language did the AI use. I would say you should not read too much into the code. It should be treated just like any other tool that you install from a package manager. We usually overlook what those packages contain, even for sensitive matters such as security vulnerabilities. We have faith that those programs are usable and secured enough.