No any instruction, rule but code-based inference effectively for AI agent

I'm a backend developer using AI agent. Recently, I faced the same problem repeatedly. The problem was wrong outputs of AI agent which ignored my instruction.

I didn't mention a change log in my Markdown Skill but the AI Agent inserted that in new document weirdly.

The reason why the AI agent did like that had been clear in the end. The AI agent actually had been following the style documents in the same directory. Since I did not say to the AI agent don't follow the others, it did own way.

And so there were 2 ways to make it do like I want. the first way was adding rules and forbid it. the second way was removing change logs in the other documents. I chose the later one because, I think it is more efficient then building a restriction instruction.

---

We face unexpected behaviors of AI agent often. Adding instructions to forbid it is almost effective. But the rules will be growing up along as code-base will be. Too much rules may cause of confusing for AI agent.

Less rules, more efficient

The behavior of AI agent is simple. If you provide it with details, it will follow them. If you don't, it will do by oneself with its mechanism. But no detail don't lead it to a quite Random. Its basic harness is going to find and read related references then it works following them by its inference.

Even if you provide no instructions but references in code-base included accumulated conventions. Though you may not instruct but code will. So you can use this mechanism instead of specification.

Using AI agent basic behavior, you can focus into the structure of code-base and it's form. This approach has a long term benefit.