I agree with you at this point. Even though Google is performing well on benchmarks and releasing impressive models like World Models Genie 3, the Gemini CLI suggestions/changes feel overly formulaic. Almost like its priorities are that of an OCD coder that cares more about tabs vs spaces instead of building a useful feature. For example, in a recent project, Google CLI spent all of my token allotment for that day on trivial tasks like tweaking ESLint configs or modularizing code that didn't need modularization.
In contrast, Claude Code seems to interpret my prompts better and helps me ship real product features for users.
Maybe it’s a system prompt issue. Its likely my prompting causing the problem. But Claude Code seems to understand my intent better.
It's how these models/their-harnesses (e.g. the Claude Code js program) are being trained together in the RL stages.
I think the software is now a very important part of the training process. Which is why I think frontier labs are only capable of shipping "actual" agents.
Anthropic has figured something out here that others have not.
I’m being down voted. I don’t have an agenda. I’m simply sharing my experience. If you’re getting good results with Gemini CLI as an alternative to Claude Code, please let me know what you’re doing to get that performance.
I’m impressed by Gemini Pro 2.5’s NLP capabilities. I use that model in production on several projects. My comments are directed only at Gemini CLI. Which FWIW is better than OpenAI Codex CLI, but much worse (for me) than Claude Code.
Even with Pro, the strict token limits combined with the model's tendency to add unrequested modifications means I run out of tokens before completing my intended tasks. Others have the same issue
https://github.com/google-gemini/gemini-cli/issues/4300
Perhaps this is the modern version of "every company ships its own org chart"? Maybe Gemini's priorities are those of a Google engineer, Claude's are those of an engineer at Anthropic....
In contrast, Claude Code seems to interpret my prompts better and helps me ship real product features for users.
Maybe it’s a system prompt issue. Its likely my prompting causing the problem. But Claude Code seems to understand my intent better.