Stop trying to automate away your own brain

Most of us use advanced technology every day without having any idea what is actually going on under the hood. We send emails in Gmail, search for information in Chrome and brainstorm ideas with Gemini without ever thinking about the source code, the architecture or the algorithms behind them. We know what we want to achieve, while the tools handle the technical execution.

It is exactly the same shift that is now happening in software development, thanks to vibe coding. For a novice or non-coder, tools like Claude Code open doors to a world that used to require years of software development. You don’t need to write syntax or understand complex class structures to build working applications.

But when the goal is to automate your own manual work, a fundamental question quickly arises: where, exactly, does the line go between what the machine should do and what the human must keep control of?

When I started developing my project ALTRUIXM, the original hope was to build a fully automated content pipeline. The goal was a self-sustaining factory where the AI would do the research, write the material, evaluate itself and continuously enrich its own database to make future searches even smarter.

It quickly turned out, however, that the road from theoretical automation to practically useful tools has less to do with software engineering and a great deal more to do with strategic domain logic.

The whole journey took off during one intense day in late February, when the first experiment, the content pipeline, came into being. The idea was simple: to prove that a model could handle the entire chain from research to finished text. It worked just as intended, and the project was then put on ice for a while as other priorities took over during the spring.

A few weeks later I took the next step from a completely different angle and built a psychosis app on an entirely new tech stack. If the first project was about producing content, the second was about mapping where the gaps in the workflow were, and deliberately stopping at an approved requirements specification (PRD).

The big breakthrough came on 19 March, when the two halves were tied together in a shared architecture. Rather than two separate apps happening to fit together, each part was built to test its own specific part of the calculation.

In the new product definition, the theory was fixed: the process needed to be split into an A-process for analysis and research and an E-process for execution and creation. But most importantly, there was the human checkpoint in between.

The first app was redefined shortly afterwards to become a component donor to the combined system, and during one intense week in July the project went through 63 commits to build the necessary guardrails that stopped the model from marking its own homework. While Fable 5 was still available, it was time well invested.

But it was once the system was up and running that the most important lesson revealed itself. In the pursuit of total automation, I had built in a function where generated text was automatically fed back into the source archive. The result was what in AI circles is known as Habsburg AI. When the model started consuming its own material, it was in effect eating its own tail. The content inbred its own phrasings and ideas, which over time led to degraded quality. Combined with the difficulty of fully trusting the semantic mapping and the settings of the vector database, the conclusion was crystal clear. The automatic addition of AI-generated content was stopped entirely.

ALTRUIXM ended up as a fantastic tool, but not as a fully automatic factory that replaced my brain or my ability to see complex connections. Instead it became an invaluable co-pilot. The system takes care of the heavy groundwork, structures the source material and makes sure the right brand voice is kept intact over time.

In the end, vibe coding isn’t about building the human out of the equation or pretending to be a software engineer. It’s about shifting roles to become the architect who sets the rules of the game. The AI model can do the heavy manual work and hold the structure, but it is still you who must provide the judgement, the strategy and the final quality control.