Pro Tip: Use GEMINI_SYSTEM_MD for better quality responses from Gemini 3 Pro Preview

[guidedways]

I've been baffled as to why Gemini 3 Pro Preview seems to respond better in Gemini AI Studio and via direct API calls (with custom instructions) but performs so, so poorly within Gemini CLI. I've tried to use Gemini for months (starting with 2.5 Pro) but it's been the same every time.

It turns out it's the built-in system prompt when GEMINI_SYSTEM_MD is not used! After going through the system prompt it seems it's geared towards vibe coding + web stacks. It's got such mixed signals all over the place, e.g.

- **Libraries/Frameworks:** NEVER assume a library/framework is available or appropriate ...

Immediately throws Gemini off for a large and complex Obj-C + Swift project with dozens of modules and frameworks. There are so many biases in the system prompt that no wonder I've been baffled why some people are able to use it and are thrilled but every time I try, my code turns into spaghetti and Gemini begins writing bogus unit tests and falling over itself.

I experimented with writing my own system prompt geared towards my workflows and use-cases and wow, the difference is night and day. There are still issues here and there but I'm seeing a huge difference in the what I've been able to achieve in one day.

So in short - if you're struggling with Gemini CLI and cannot seem to make it work for you, try overriding GEMINI_SYSTEM_MD and point it to a custom system prompt. You should be able to steer it a lot more accurately.

Mini tutorial:

1. Create profiles (i.e. more than one system prompt) if you manage multiple projects. Here's how I created one when working on Obj-C/Swift projects:

mkdir -p ~/.gemini/SYSTEM_PROMPTS
nano ~/.gemini/SYSTEM_PROMPTS/profile_swift.md

Add your custom instructions to the file.

2. Create shell functions in ~/.zshrc (or ~/.bashrc)

# Helper function to run gemini with a custom system prompt
_gemini_with_prompt() {
  local prompt_file="$1"
  shift 1

  if [ ! -f "$prompt_file" ]; then
    echo "gemini: missing system prompt $prompt_file" >&2
    return 1
  fi

  GEMINI_SYSTEM_MD="$prompt_file" gemini "$@"
}

# Swift development profile
gemini-swift() {
  _gemini_with_prompt "$HOME/.gemini/SYSTEM_PROMPTS/profile_swift.md" "$@"
}

# Bug hunting profile
gemini-bug() {
  _gemini_with_prompt "$HOME/.gemini/SYSTEM_PROMPTS/profile_bug.md" "$@"
}

3. Reload your shell

source ~/.zshrc

4. Use it

# Run with Swift development system prompt
gemini-swift

# Run with bug hunting system prompt
gemini-bug

# Or use it inline for a one-off:
GEMINI_SYSTEM_MD=~/.gemini/SYSTEM_PROMPTS/my_prompt.md gemini

---

The key is the GEMINI_SYSTEM_MD environment variable; set it to the path of your custom system prompt markdown file before invoking gemini.

[jackwotherspoon]

This feels more like a great discussion post and less of an issue per-say.

[guidedways]

@jackwotherspoon indeed, although I wanted to highlight that the default system prompt overly leans towards a certain type of development setup. This is why it's never worked for our setup (and I know many developers personally that have tried and had similarly experiences and moved on).

I am hoping the default prompt can be given a second glance and revised and potentially made more generic.

I've compared this with the default prompt from Codex and they use a super generic set of instructions which is why it works wonders out of the box with zero tweaks.

[guidedways]

To add to this, the current system prompt in the CLI goes against the new set of instructions on how to prompt Gemini 3 in particular. It's doing quite the opposite - the current system prompt is being very specific, referring to specific stacks:

  - **Full-stack:** Next.js (React/Node.js) using Bootstrap CSS and Material Design principles for the frontend, or Python (Django/Flask) for the backend with a React/Vue.js/Angular frontend styled with Bootstrap CSS and Material Design principles.
  - **CLIs:** Python or Go.

This just confuses the model when placed inside of a C / C++ / Obj-C environment with crazy amount of custom frameworks where we'll never be doing any web development or design work etc.

[guidedways]

@jackwotherspoon I've created a separate bug report: #13787

[jackwotherspoon]

Thanks @guidedways i think we are going to be making some more default system prompt changes shortly. So will definitely flag the new issue to the team.

[guidedways]

Thank you, this would most certainly help.

[yaohsien-md]

I've been baffled as to why Gemini 3 Pro Preview seems to respond better in Gemini AI Studio and via direct API calls (with custom instructions) but performs so, so poorly within Gemini CLI. I've tried to use Gemini for months (starting with 2.5 Pro) but it's been the same every time.

It turns out it's the built-in system prompt when GEMINI_SYSTEM_MD is not used! After going through the system prompt it seems it's geared towards vibe coding + web stacks. It's got such mixed signals all over the place, e.g.

- **Libraries/Frameworks:** NEVER assume a library/framework is available or appropriate ...

Immediately throws Gemini off for a large and complex Obj-C + Swift project with dozens of modules and frameworks. There are so many biases in the system prompt that no wonder I've been baffled why some people are able to use it and are thrilled but every time I try, my code turns into spaghetti and Gemini begins writing bogus unit tests and falling over itself.

I experimented with writing my own system prompt geared towards my workflows and use-cases and wow, the difference is night and day. There are still issues here and there but I'm seeing a huge difference in the what I've been able to achieve in one day.

So in short - if you're struggling with Gemini CLI and cannot seem to make it work for you, try overriding GEMINI_SYSTEM_MD and point it to a custom system prompt. You should be able to steer it a lot more accurately.

Mini tutorial:

1. Create profiles (i.e. more than one system prompt) if you manage multiple projects. Here's how I created one when working on Obj-C/Swift projects:

mkdir -p ~/.gemini/SYSTEM_PROMPTS
nano ~/.gemini/SYSTEM_PROMPTS/profile_swift.md

Add your custom instructions to the file.

2. Create shell functions in ~/.zshrc (or ~/.bashrc)

# Helper function to run gemini with a custom system prompt
_gemini_with_prompt() {
  local prompt_file="$1"
  shift 1

  if [ ! -f "$prompt_file" ]; then
    echo "gemini: missing system prompt $prompt_file" >&2
    return 1
  fi

  GEMINI_SYSTEM_MD="$prompt_file" gemini "$@"
}

# Swift development profile
gemini-swift() {
  _gemini_with_prompt "$HOME/.gemini/SYSTEM_PROMPTS/profile_swift.md" "$@"
}

# Bug hunting profile
gemini-bug() {
  _gemini_with_prompt "$HOME/.gemini/SYSTEM_PROMPTS/profile_bug.md" "$@"
}

3. Reload your shell

source ~/.zshrc

4. Use it

# Run with Swift development system prompt
gemini-swift

# Run with bug hunting system prompt
gemini-bug

# Or use it inline for a one-off:
GEMINI_SYSTEM_MD=~/.gemini/SYSTEM_PROMPTS/my_prompt.md gemini

---

The key is the GEMINI_SYSTEM_MD environment variable; set it to the path of your custom system prompt markdown file before invoking gemini.

I had the same experience with Gemini CLI which made me so sad. I am glad to know the reason Now. Thank you very much for sharing this essential information.

[guidedways]

Essentially I've been complaining (and trying) for several months to see if gemini cli can perform as well as other CLIs in the space, but so far it's been very disappointing.

I love Gemini as a model. You can see my bias with this open source MCP server I developed, that lets you connect any CLI with any AI model, primarily recommending Gemini: https://github.com/BeehiveInnovations/zen-mcp-server

However, when you use Gemini CLI directly, it feels very different from the one you experience via the API. So I started poking around the code to see if it's in fact the system prompt causing this behavior (since with Zen MCP I use the Gemini 3 Pro recommended system prompt with a well defined generic role and a set of generic instructions) and yes - the system prompt in fact was causing a lot of the unwanted behavior.

The one I replaced it with now forces Gemini to think harder, perform logical reasoning steps and so on, and is generic enough that it works a lot better.

For perspective, every single time I've tried to use gemini for anything, I've given up after two prompts (in the past several months, I have been testing GeminiCLI at least every second day).

Only after updating the system prompt was I actually able to use it continuously for a couple of hours and solve a few problems. For the first time ever, Gemini beat Codex in finding and fixing a concurrency related bug.

[binarypie]

I'm going to give this a shot. I've been switching between Augment Code (Claude Sonnet 4.5) and Claude Code for a while to actually help with getting work done. However, I find that Gemini generally has better architectural sense. I'm often designing in Gemini and implementing in some version of Claude.

I'd like to get to the point where my expense of agents costs can be somewhat predictable because jumping around every 30 days is becoming more annoying than writing Javascript full time during the Web 2.0 days.

[guidedways]

Gemini is incredibly strong in reasoning, but it seems to overshoot 75% of the times and is incredibly unpredictable thus far. I've had additional success by overriding modelConfigs and customAliases in settings.conf by effectively overriding the base gemini-3-pro-preview with custom temperature and topP settings. I think the default of 1 (as Gemini 3 docs recommend) is suitable for most work but probably not code - it's goes into a thinking loop every 3rd prompt.

As much as I'd want to daily gemini-cli, codex seems to work best for my use case at the moment. Claude Code is excellent when it comes to refactoring code (once you have a plan laid out by Gemini / Codex) but entirely useless for anything else that involves reasoning, architecture, good structured code etc.

[binarypie]

Yeah I've been having long running web chat conversations with Gemini Pro to come up with an architecture. I turn that into a set of markdown files that create the boundaries for each step of implementation. Then I use Augment Code to take the first pass because it generally does a better job at large steps. Then I layer in Claude Code for refactoring. Finally I use Gemini again at the PR side to do a code review with Claude to fix bugs. After that. I manually read all the code, do QA, and write more integration and unit tests. Fix the last bugs then merge. It's pretty interesting as I very much feel like I have a team of 10 people working with me at any give moment but they are all about mid-level in experience and skill.

[guidedways]

That's very similar to my approach (albeit with codex in the mix) and one that works remarkably well with any tech stack. Doing it this way is probably the only sane way to use AI agents in the mix without ending up with regressions and new and scary bugs, finishing with a self-QA and an exhaustive test suite to keep you covered.

And as you discover bugs, TDD all the way; write a failing regression test, get your favorite model to fix the bug, confirm it's fixed, review, look for additional edge cases and cover those, commit and merge.

[davidsgrogan]

How did you find the content of the built-in system prompt?

[yaohsien-md]

How did you find the content of the built-in system prompt?

I have the same question. Please help us find out the built-in system prompt. Thank you very much.

[andromarcestf]

The built in system prompt is here: https://github.com/google-gemini/gemini-cli/blob/main/packages/core/src/core/prompts.ts

@davidsgrogan @GodBlessingMe