Gemini CLI stubbornly insisted its code was right and even did a Google search to prove it with an outdated article.

I was speechless. 😂

One of the main problems with AI coding agents now is that they’re still not fully reliable in following instructions.

Based on my experience, it’s about an 80% hit and 20% miss.

So, we need to read the code to ensure it follows our instructions. If not, adding follow-up prompts or resetting the changes and retrying sometimes works.

But sometimes, it feels like working with a stubborn software engineer who refuses to acknowledge they’re wrong.

In my recent case, Gemini CLI ironically generated old and non-working code for its own Gemini API.

What made it very annoying was that it refused to refer to the latest code samples I provided from the official documentation, and insisted that its own version was the most up-to-date.

Even more annoying, when I insisted my code was the latest, it pulled an old reference from Google and still claimed its version was the correct one.

Well, I had to bite the bullet and fix it myself. 🤷‍♂️

It’s a reminder that AI is a tool to augment our abilities, not a replacement for our expertise. We still need to be the final reviewer and the one in control.

What are your experiences with AI coding agents? What are your thoughts?

—

I share practical tips about business, marketing and AI 🔔 Follow me to learn more! ♻️ Re-post this to help others! 🔖 Save this for future reference! 💬 DM me for collaboration!

#GenAI #Coding #SoftwareDevelopment #Tech #VibeCoding