Use AI to Get Rid of AI

Warning: This text is 100% human written (mistakes and all)

As I wrote a few days ago, I have been a heavy user of AI over the last months. I have spent a lot of time in Claude Code and the terminal (my natural habitat — it’s fun, that it makes a come back (after too many years spent in Powerpoint, Confluence and Jira) and I enjoy using it (mostly — the CC application has some warts, but that’s another story).

I am sure, you all have experienced the thrill when your work with AI starts to be mindblowing — because the stochastic word generators that they are — produce something that you didn’t think they could.

But surely you also have experienced the frustrating moments, when the LLMs just make up stuff. Or have some really loose interpretation of your prompts. (My favorite one: I have very strict rules for software development with tooling that writes specs, plans, tasks, tests, and then implements, and reviews and verifies. Often I saw the LLM „reason" (and this is in quotes, because they really aren’t reasoning, are they) that the task given was so simple that it could safely skip the whole planning ceremony and go directly to implementation).

So I am working on systems that take away the autonomy of the LLM. Where the flow is controlled by code, without any LLM interaction. I let the AI write software that takes away its agency. That yields much better results in the long run. AI is not a tool for everything (and let’s not even touch the data protection implications) and using it as the hammer for all problems is just wrong on so many levels (cost, energy cost, time cost, token cost, non-determinism). But using it to write specialized tools — that’s a totally different thing.

I have dozens of small command line tools that do one thing and one thing only. They are mostly composable — the output of one becomes the input of the next. I have workflow engines that compose these actions into flows. Some of the actions use LLMs (if they really need it).

This has so many advantages:

• fast
• cheap
• correct
• repeatable
• privacy respecting (the AI never sees the data — or if it does, I can filter it before it goes over the wire)

This is the way, I think, we must work towards. Just using AI as a hammer seems similar to using a wrecking ball to drive a nail. It works, but the side effects are probably not what you wanted.

Your thoughts?

Originally posted on LinkedIn.