Wrote my first LLM Skill

For a significant period of time I was pushing it, providing a variety of reasons(mainly because I’m old school) in the end I gave it a shot starting with Claude Code in my work process in the beginning it was fun as it was something new for me, started small with some FE tasks a few researches and it was fine step by step It started to be a bigger part of my workflow.

After a while I noticed that I was getting more tired of coding with it the LLM than by myself, it’s a known risk and it’s well documented.

But then I was introduced to Matt Pocock and everything I knew about the use of LLMs changed in a instant.

Why should you let it lead you when it needs to be the other way around, that’s why I started to look a bit different on the whole thing.

I started to look into what skills Matt Pocock has and start to digging deeper in the way skills are created, defined, used and then I realized that there was something that happen to me quite often when using Claude Code, when I ask something either a question or something to be done it does it for memory(training data) does not look into the code or any documentation…

I've had a few cases, when I simple feature that was made with LLM had missing essential things. One vivid example that comes to my mind is with Str::slug method, I needed simple string operation to be done and the LLM decided it will be better to write almost the whole method from scratch...and even when I pointed out that is not necessary it insisted I need to provide the path to the class itself to "convince" it that I'm right...and then there I've decided that I need to write a skill to "fight" that behavior.

That’s why I’ve created my own skill: research-before-answer

It’s not anything fancy it’s a simple:

When this skill is active, you MUST NOT answer from memory. Every claim about a package, framework feature, method signature, or behavior must come from an authoritative source consulted in this turn.

it forces the LLM to do a research before saying anything stupid or outdated, of course they will be cases where this skill won’t work but the general idea is make it a bit “smarter” and not give a quick answer that can cause problems in the future.

Conclusion

My main goal working with LLMs is: Guide them, don’t let them guide you there a lot of nice things using it, but we need to be careful how we use it, don’t let you dev skills became opsolite, used it and improve it with the help with the LLMs. The future is now