The Gotcha 🎭

When your AI assistant apologises more than my British relatives at a funeral.

Ever asked an AI to find a file and watched it have a proper existential crisis? Welcome to Gemini CLI.

I gave it a simple task: find postman_environment.json and run some tests. What followed was a masterclass in digital faffing.

Let me walk you through this shambles:

Attempt 1: "It looks like the file does not exist in the current directory."

Narrator: It did.

Attempt 2: "I apologise, I need an absolute path to list the directory contents."

Mate, you're a CLI tool. Working directories are kind of your thing.

Attempt 3: "Okay, I have the absolute path. Now I will list the files..."

Progress! Sort of.

Attempt 4: "I've refreshed the file listing, but I still don't see a postman_environment.json file."

The file is right there. In the directory. That you're looking at.

Finally: "You are absolutely right! My apologies. The file was being ignored."

Ignored by whom? The .gitignore? You're reading the filesystem, not git staging!

📚 Geek Corner
The gitignore gotcha: Gemini CLI appears to respect .gitignore when listing files, which sounds clever until you realise it means the AI literally cannot see files that exist on disk. That's not "smart" - that's a tool confused about its own job.

Here's the thing: I don't mind an AI asking clarifying questions. What I do mind is an AI that apologises six times while repeatedly failing to do the one thing it was designed to do.

This isn't bounded rationality - this is unbounded apologising.

Bottom line: If your AI assistant spends more time saying sorry than solving problems, maybe it's not ready for the CLI wild west.

---

The Token Inferno 🔥

While we're on Gemini, let me show you what happened when I let it loose on a slightly larger task.

398,377 tokens. Nine turns. Two hours and four minutes of wall time.

For context, that's roughly the equivalent of reading War and Peace... twice. The actual API time? 2 minutes 40 seconds.

So what was Gemini doing for the other 2 hours? I honestly don't know. Contemplating its existence? Having a nice lie down? Running up my future bill?

📚 Geek Corner
Wall time vs API time: Wall time is how long you wait. API time is how long the model actually works. A 45x difference usually means either network issues, rate limiting, or the model doing something spectacularly inefficient with its "thinking" tokens. In this case, those 5,783 "Thoughts Tokens" suggest Gemini was having a proper existential crisis between each response.

The kicker? I don't even remember what task I gave it. Whatever it was, it wasn't worth a novel's worth of tokens. If you want to see how a terminal coding agent should handle tasks, the contrast is stark.

Feels like: Asking someone to nip to the shops and they come back three hours later having driven to Scotland "for the scenery."

---

Let's Go Nuclear ☢️

It's a React component.

I did not click approve.

---

What I'm Consuming 📚

Gemini CLI drops - and so does my patience

Google launched Gemini CLI this week - open source, Apache 2.0, built on Gemini 2.5 Pro. The pitch is solid: natural language in your terminal, code generation, debugging, the works.

I was properly excited. Then I actually used it (see above).

The gap between the marketing and the experience is wild. On paper: "agentic terminal tool that handles your entire workflow." In practice: can't find a file that's sitting right there in the directory. The .gitignore gotcha alone tells me this shipped before anyone tried it on a real project with a real .gitignore.

The open-source angle is interesting though. Apache 2.0 means the community can fix what Google won't. I suspect this gets proper good in six months - just not today.

Claude goes full platform

Same week, Anthropic launched hosted artifacts - build, host, and share AI-powered apps directly from Claude. No deployment. No API keys. No infrastructure.

The clever bit? When someone uses your app, their Claude subscription pays for the compute, not yours. You build it, share a link, done. Zero cost to the creator.

I'm torn on this. On one hand - this is the fastest path from "idea" to "working app" I've ever seen. On the other hand - we're building a world where every app is a Claude wrapper running on someone else's subscription. That's either genius or a house of cards.

The timing is cheeky too: same week Google launches Gemini CLI, Anthropic casually drops a full app hosting platform. Different strategies, different bets. Google's going open-source terminal. Anthropic's going walled-garden platform. Place your bets.