This weekend I put Claude Code vs Gemini CLI head to head in a yaml wrangling contest.

Gemini CLI connected to gemini-2.5-pro Claude Code subscription - Pro

Results - Gemini CLI PR - Claude Code

Both are designed and positioned as *coding assistants. Yaml is not necessarily coding, but this is how I make my living and my area of expertise.

I gave both contestants the same content in GEMINI.md/CLAUDE.md and the same very detailed task description in task-.md. The tasks are identical except I added instruction for Claude to add a suffix to names to avoid collisions in shared project. They executed on separate branches without ability to see each others code.

The task was to create a Crossplane Cloud Run Composition with required accompanying resources like IAM. The service itself should implement invoker check, Direct VPC Egress, b/g deployment.

Claude diligently read all the documentation I pointed it to. It spend nearly 36k tokens on it. This is my fault because I included generated content (zz_generated...) which is not only duplicated, but also very voluminous and confusing noise. But I didn't notice gemini spending too much time reading it

Not tested yet, but Claude seems more on mark than Gemini even though gemini is a google house native. gemini totally missed Direct VPC Egress. Even after I pointed it out, its re-take wasn't accurate

TODO

Gemini CLI box dances around the screen a bit, not too dramatic but not as clean experience as Claude Code. I like token summary (execution time is skewed because I stopped it following morning)