Parts of Gemini CLI's system prompt look eerily familiar

I've been playing around with the idea of building my own commandline coding agent, both to better understand how these things work and to try out a couple of things that I haven't seen in any of the existing tools.

A couple of the things I'm very interested in are how these agents tools get defined and described to models and what their internal 'system prompts' look like.

Anthropic is famously diligent in how much effort they put into their system prompts and model instructions. Claude Code is distributed as obfuscated JavaScript. The system prompt is baked into the source code in plain text, so you can see if by inspecting the 7MB+ 'claude' file. You can also see it if you inspect the HTTPS traffic between Claude Code and Anthropic's API servers. But, to the best of my knowledge, the prompt isn't...freely available. It's part of a blob of code that's marked:

OpenAI's codex-cli is open source, but a lot lighter on the system prompt than Claude Code.

I was really excited when Google shipped their gemini CLI tool a couple days ago. It's open source (Apache licensed) and has...exactly the kind of detailed prompting that Anthropic is so good at.

As I started reading through some of the prompts baked into gemini, they seemed...familiar.

I was a little surprised at just how familiar the prompting seemed, when I compared it to Claude Code's prompts.

This is Claude Code:

claude verbosity

This is Gemini:

gemini verbosity

They're clearly not the same text.

But it doesn't feel like a coincidence that the first few examples of "good" responses in Google's Gemini coding tool, look a whole lot like Anthropic's.

Anthropic:

<example>
user: is 11 a prime number?
assistant: Yes
</example>

Google:

<example>
user: is 13 a prime number?
model: true
</example>

At least to me, it sure looks like what's going on is that someone asked a model to "wash" Anthropic's prompting.