Highly Positive Feedback with a Key Bug Report: Gemini CLI Excels Over Claude Code

[JamshidShakirov]

Dear Gemini CLI Development Team,

I'm writing to share extremely positive feedback regarding my experience with Gemini CLI, especially when compared to Claude Code. I've found Gemini CLI to be a significantly more effective and professional tool for code generation and management, particularly when adhering to technical documentation.

Here are the key areas where Gemini CLI truly stands out in my workflow:

Conceptual Integrity and Contextual Awareness: Gemini CLI consistently demonstrates a remarkable ability to maintain conceptual integrity within a project. It constantly checks the overall context of my existing codebase, ensuring that new code integrates seamlessly with the architecture and logic. This is crucial for large projects.

Duplicate File Prevention: A major benefit I've observed is its proactive approach to preventing duplicate code files. This feature helps maintain a clean, efficient, and well-organized codebase, saving significant refactoring time.

Superior Pull Request (PR) Quality: The code generated by Gemini CLI leads to higher quality Pull Requests. Its outputs are well-structured and align effectively with project standards, which streamlines the code review process and significantly enhances our development workflow.

Professionalism and Human-like Interaction: This is perhaps the most impressive aspect. Gemini CLI genuinely outperforms Claude Code in its professional approach to tasks. It delves deep into the problem's essence when generating code. Furthermore, the interaction feels incredibly natural; it's like communicating with a human, not just an AI, due to its ability to understand nuances and respond thoughtfully.

Critical Bug Report: Task Looping
Despite these excellent features, I've encountered one critical bug that impacts efficiency:

When Gemini CLI is presented with a task, it sometimes reverts to a previously completed task and attempts to solve it again, even though it has already been addressed. This loop wastes time and requires manual intervention to steer the CLI back to the current, unaddressed problem. Addressing this would further enhance the user experience and productivity.

Overall, Gemini CLI has become an indispensable tool in my development process. Congratulations on creating such a powerful and intuitive product, and I hope the bug report helps in further refinement.

Thank you for your excellent work!

[scidomino]

Glad to hear it's working well for you!

Mitigating looping is high on our list of things to fix (we discussed it just this morning). So hopefully you should see some progress on that in the coming weeks.

Reverting is also a known issue we are working on.

[HanymiGame]

I used this client with an API key from a Tier 1 account, and in an hour of work on a not very complex project, this client generated 20 million input tokens for me, well, in general, the account came to about 10 dollars. I think that the context is not optimized there. It increases it with each request, although it would be great to summarize the information automatically, I did it manually /compress, but in automatic work it generates a huge number of input tokens. In his place, I would collect general information about the project in one file so as not to make unnecessary requests and refer to this context, add and delete current data from it. So far it looks economically unprofitable, compared to a regular chat. On a free account, when the request limit was reached, it simply hung on an error and that's it, and continued to stubbornly make requests. But I couldn't even change the model, I had to exit and lose all the progress.

[keithballinger]

@HanymiGame We definitely don't want this outcome. (Even I with a personal key have ran into some of this recently.) It's very, very high on our priority list to be best in class with efficiency of token mgmt. Not the best answer in the meantime, but one thing I've been doing is creating a file with my technical design (usually I call it arch.md), and then a task list in plan.md. I put general insructions in plan.md above the tasks to instruct the agent to mark the task done in plan.md and commit. And then, I'll often /clear after each task to keep the context clean. (I originally started doing this with agents even before we created gemini cli, because I find it super useful to help any agent continue to stay focused.)

[HanymiGame]

Yes, that's a good point from you. I do the following to optimize the context with a regular chat. Let's say you have a project, write a simple script that recursively goes through all the project directories and collects all the sources into one flat file with the path and compressed source code inside, in MD format. The model doesn't care how you provide the context, it doesn't need comments in the code, it understands everything itself, you just maximize, compress the code into one line, removing everything unnecessary and unimportant. You give this context in the first message in the chat and then refer to it, it all works. This approach is not quite suitable for this tool, since this context will probably not be cached. I would rather use a combination of your and my solutions, then the model will not need to physically read any files during subsequent requests, it will already know the entire project from the context and operate directly with the files on the disk - without searching.

[Manamama]

Yes, I use a lot of "memory jog" .md files, also for myself. Part of PM tracking anyway...