Gemini cli seems incapable of *not* writing code

[tylertronn]

I've tried adding very clear instructions in the root GEMINI.md file, stating to never write code/vibe code for me, giving the agent a persona of a mentor/teacher, outlining the goals of the cli are purely to help me learn/plan/research, but alas, nothing changes.

I do not want to vibe code - I enjoy programming, but I would love to use the tool purely to help me plan projects, help me research libraries or tools, give me advise on implementation - act like a coach and an actual assistant, rather than "do my job for me"

No matter what, it always seems to slip up and when called out on, it realises it. What can I do to help ensure the agent knows it shouldn't just write my code for me?

[okurz]

You should be able to use the new /plan command in version 0.29.0 . Alternatively I have already in the past written a custom command /plan in ~/.gemini/commands/plan.toml with the following content

description = "Create a plan in the tasks/ directory without triggering any remove or add Git actions"
prompt = """
Create a detailed execution plan for the following request: {{args}}

SAVE this plan as a markdown file in the 'tasks/' directory.

STRICT RULES:
1. Do NOT stage (git add) the new file.
2. Do NOT commit the file.
3. Only perform safe file operations, only read-only 'git' operations.
"""

[mystilleef]

Try looking at this project to help you setup the CLI properly.

https://github.com/mystilleef/gemini-cli-modes

The setup in that project brings it close to Claude Code for me, if you use the modes.

I've also heard good things about the conductor extension.

[aniruddhaadak80]

From my point of view, the gap here is usually that the default agent loop is optimized to be helpful by doing work, while your intent is more mentor style guidance. Plan mode and custom commands help because they change the execution path, not only the wording of the instructions. I would also treat this as a place where the product could expose a clearer coaching mode instead of expecting users to simulate it only through prompt files.