Discussion
I read this blog post from Jacob Rockowitz and kept on nodding my head in agreement. As LLMs get better and better, I find that agent skills are often just bloating my context window. 🤯📚
https://www.jrockowitz.com/blog/vibing-drupal-losing-my-skills
Also - yes, I do fear that my (brain) skills could atrophy by using coding agents. I am trying to mitigate this by not allowing the agent the ability to make any commits - I've been manually reviewing and committing all changes - so far... 🧠⚠️
> the nagging feeling every kid has taking off training wheels: is my own balance getting better, or am I just relying on a better bike?
This metaphor from Claude at the end is unintentionally perfect - I doubt any kid has this nagging feeling, let alone 'every kid'. Most are just thinking 'wow I can cycle without stabilisers now!'.
That kind of confident statement - that scans well, that supports the theme of the blog, but appears false - is precisely why I don't trust LLMs in critical work. If bloggers like Jacob don't notice this, how to trust they'd spot similar flaws in their LLM output that still reads well and supports the goals of the prompts?
Install bonfire.cafe
Get the full app experience
