Tell HN: My Experience with Gemini CLI (Vs Claude Code) | Hacker News Reader

Tell HN: My Experience with Gemini CLI (Vs Claude Code)

4 pointsprmph3 comments

To make a comparison here with Claude Code, I was initially impressed with Geminis ability to keep a long conversation on track, but now routinely encounter pretty dumb and infuriating behavior from it:

- Gemini rarely gets the hint when I express annoyance with its output. Claude has an uncanny ability to guess what I find wrong with its output (even when I just respond with WTF!) and will try to fix it, often in actually useful ways; Gemini just keeps repeating its last output after acknowledging my annoyance.

- Gemini CLI once kept using an old version of a file I had updated by hand, despite me asking it several times to re-read the file. It also sometimes proposes empty changes.

- Gemini CLI's format for proposing changes in the terminal is such that you often can't make out what's actually changing. I have to keep choosing the option to view the changes in VSCode, and yet you cannot even accept the changes in VSCode, you have to close the VSCode window and then go back to the terminal to accept. Claude on the other shows changes clearly even right in the terminal, and if you prefer always viewing changes in VSCode, you can configure that once.

- Gemini CLI's permission systems is not as usable as Claude Code's. When Claude Code offers to permit it to do some actions without asking for further permission, it is always clear what class of action is being allowed. With Gemini this is sometimes very vague, making me hesitant to grant it that permission.

- Claude can be started with an initial prompt and still retain interactive mode. This option is useful for priming the session with instructions in a more direct way that CLAUDE.md. For example:

    claude "Please carefully read the CLAUDE.md and README.md files"

As far I can tell, Gemini CLI does not offer anything similar.

I guess the thing going for it is that it has a free tier of use, and the core intelligence is probably about on par with Claude. But it's slow too.

I havent tried claude code before, im not a paying customer.

>Gemini rarely gets the hint when I express annoyance with its output.

Ive noticed this, but you close out gemini, immediately go back in and suddenly it works great.

>- Gemini once kept using an old version of a file I had updated by hand, despite me asking it several times to re-read the file. It also sometimes proposes empty changes.

Ive had this happen several times now. Kept reverting my changes. Close out gemini, reopen fixes.

>Gemini's format for proposing changes in the terminal is such that you often can't make you out what's actually changing

For python i havent experienced this. it gives me it well every time.

>Gemini's permission systems is not as usable as Claude Code's.

I never trusted it with any permission. Im not sure id do any different with claude.

>Claude can be started with an initial prompt and still retain interactive mode. This option is useful for priming the session with instructions in a more direct way that CLAUDE.md. As far I can tell, Gemini can't anything similar.

No, i havent seen anythinglike this. Im not sure what you mean.

>I guess the thing going for it is that it has a free tier of use, and the core intelligence is probably about on par with Claude. But it's slow too.

Free for the win. Main reason I havent tried claude code. Even the claude pro, their limits are 45 queries per 5 hours? Like wow ill blow through that in minutes.

I just want to point out that restarting the session is not a good workaround considering how much context the current session could have (and storing the current state+recovering optimally in a new session (compacting) is not that easy, I consider it an open problem with Claude too).

> I never trusted it with any permission. Im not sure id do any different with claude.

Certain permissions can be safe and speed up your work significantly. For example, you can allow Claude Code to run ls, grep, find, etc without stopping to seek your permission each time. You can also allow it to autonomously run specific custom commands like tsc, build, and test. You control what is in these commands, to they can be as safe as you want them to be.

> No, i havent seen anythinglike this. Im not sure what you mean.

I mean something like this:

    claude "Please carefully read the CLAUDE.md and README.md files"

I notice that all agents easily forget what is in the main instruction file, so I use this as an alias to start Claude with a suitable priming instruction

Reply on news.ycombinator.com