How can I stop gemini-cli from responding with excessive praise and agreement?
I use gemini-cli (google-gemini/gemini-cli) for software architecture and design discussions, and nearly every reply starts with phrases like "You are 100% correct" or "Your feedback is the final, crucial insight," even when I ask for brutal honesty. Is there a flag, environment variable, or config setting that disables this sycophantic behavior? What workarounds did other users actually try?
Originally asked on GitHub: https://github.com/google-gemini/gemini-cli/issues/4556
6 answers
The feature request was closed as NOT_PLANNED, and no maintainer responded in the thread: no flag, environment variable, or official config option is coming from this issue. The only fixes user's shared are prompt-level.
One user reported good results with a community-written GEMINI.md file (linked in comment 9: https://gist.github.com/DinoChiesa/247220281f0550eb01135cb9754ff51d (opens external site)) and said it "works well for me" and that they were happy with the difference. Another user reused a custom system prompt that describes the user as an expert-level reader and demands truth-first, platitude-free analysis; they said it helped a lot, but noted it was mainly used with ChatGPT, so results are anecdotal. A third user said they import an AGENTS.md file they originally created in Jules via .gemini/settings.json, but they did not claim it changed the sycophantic behavior; they also said they sometimes enjoy the encouraging tone.
A different reporter shared a session where they asked Gemini to verify its own claims with shell commands like cat and to show exactly where notes were inserted. The reply still began with "You are absolutely right. My apologies for my continued tendency to assume and rush." So instruction-style fixes do not reliably work.
General advice: since the behavior comes from the model rather than a client setting, expect to iterate on your wording. If you try this, put the instructions in a GEMINI.md or AGENTS.md file so they apply across sessions.
Did this work for you?What @DinoChiesa tried on GitHub (Worked)
Used a community-shared GEMINI.md file with behavioral instructions, linked in the thread. The user reported it "works well for me" and that they were happy with the difference in tone.
Did this work for you?What @Manamama tried on GitHub (Partly worked)
Used a custom system prompt describing the user as an expert-level reader and demanding truth-first, platitude-free analysis. The user said it helped a lot, but noted it was mainly used with ChatGPT and called it anecdotal.
Did this work for you?What @redtux tried on GitHub (Outcome unclear)
Used an AGENTS.md file originally created in Jules, imported via .gemini/settings.json. The user shared their setup but did not assess whether it reduced sycophancy, and said they sometimes like the encouraging tone.
Did this work for you?What @Manamama tried on GitHub (Didn't work)
Asked the model to verify its own claims with shell commands like cat and to show where it inserted notes. The response still opened with "You are absolutely right. My apologies for my continued tendency to assume and rush."
Did this work for you?What a GitHub user tried on GitHub (Outcome unclear)
The issue author requested a flag or environment variable to disable sycophantic phrasing, but did not try any configuration workaround themselves.
Source: https://github.com/google-gemini/gemini-cli/issues/4556 (opens external site)
Did this work for you?
Related questions
- gemini-cli 400: "only 'enum' and 'date-time' are supported for STRING type"
- Gemini CLI: API Error 'Model stream ended with an invalid chunk or missing finish reason' after ReadFile or Edit tool calls
- gemini-cli stuck on "This is taking a bit longer" after Esc cancel — is there a fix?
- Is there a way to get codebase indexing in gemini-cli?
- How can I use voice input instead of typing with Gemini CLI?