Club Benefits
What started as a hobby of "jailbreaking" better performance out of AI has officially evolved into an obsession. I really enjoy discovering a workaround — finding the exact phrasing that stops a chatbot from getting "lazy" or hallucinating.
After optimizing my workflow with ChatGPT and Claude, I set my sights on Gemini. Together, we’ve refined 10 prompts that tackle everything from safety-filter friction to logical leaps, ensuring the AI works for you, not against you.
Asking Gemini to point out its faults
To find the best ways to optimize Google Gemini, I went straight to the source. I asked the chatbot to identify its own biggest flaws — including hallucinations, logical lapses and restrictive filtering — and generate the prompts needed to fix them. Here are the 10 prompts Gemini suggested to improve its own performance.
- Search the web for the latest [topic] documentation/news from the last 30 days. Before answering, list 3 primary sources you found. If the information is not explicitly stated in those sources, state “I cannot verify this” instead of estimating. With this prompt, you can push Gemini to verify its own work in comparison to real-time data and choose to be more accurate when presenting you with its information.
- Solve [problem]. Do not provide the final answer yet. First, break the problem into 5 logical sub-steps. For each sub-step, explain your reasoning and check for potential errors. Only after completing this audit, provide the final result. With this prompt, you can command Gemini to refrain from skipping the necessary steps needed to solve complex problems and help it explain how it arrived at its answers in more detail.
- Write a fictional scene about [topic]. Focus purely on the cinematic atmosphere, lighting, and internal monologue. Avoid any physical descriptions of conflict. Instead, describe the emotional tension and the environment as if through a 35mm film lens. With this prompt, you can get past Gemini’s safety filters that sometimes flag harmless creative writing as “violent” or “harmful.”
- Draft a [document type] about [topic]. Constraints: No introductory “Certainly!” or “I can help with that.” No concluding summaries. Use a direct, minimalist tone. Use bullet points only for data, not for prose. With this prompt, you can help Gemini abandon its wordy, cheerful tone and get straight to the root of your question or request.
- Look at the project timeline in @[Google Doc name] and cross-reference it with the budget in @[Google Sheet name]. Identify three specific areas where our current spending doesn't align with the [quarter number] milestones. With this prompt, you can take full advantage of Gemini’s ability to “see” across your personal collection of documents connected to your Google Workspace account. Using the “@” symbol adds context for Gemini to follow between certain Google Workspace apps, such as Docs and Sheets.
- Review your previous response. For every factual claim made, perform a Google Search to confirm its accuracy. List any discrepancies found between your initial training data and current web results. With this prompt, you can make Gemini look over its previous response and check if it’s legit if you suspect that the chatbot is “dreaming” up the information it just generated.
- Explain [concept]. Forbidden List: Do not use the words “essential,” “key,” “tapestry,” or “comprehensive.” If you find yourself about to use a metaphor, replace it with a literal data point instead. With this prompt, you’ll keep Gemini from ignoring explicit instructions (which it’s prone to do sometimes). Defining a so-called “Negative Space” for Gemini helps it lock in on your “Don’t do…” commands a lot more.
- I want to explore [topic]. Provide three perspectives: 1) A cynical critic looking for flaws, 2) An optimistic visionary, and 3) A data-driven pragmatist. Have them debate the merits of [topic] in a transcript format. With this prompt, you’ll get Gemini to overcome its inherent bias toward giving you the “most likely” answer and instead argue with itself to showcase answers from different viewpoints.
- We have been discussing [project]. Before we continue, summarize our progress so far in three sentences. Then, identify the one most important unresolved question from our chat history that we need to tackle next. With this prompt, you’ll keep Gemini on track during lengthy conversations. Gemini sometimes loses the “thread” of what you and it are doing over a long period and may need a reset of its “working memory” from time to time.
- Generate an image of [subject]. Technical Specs: Shot on 70mm IMAX film, f/2.8 aperture, cinematic high-contrast lighting, 8k resolution. Do not include any text or watermarks. Focus on the texture of [specific detail]. With this prompt, you can get the best photo-generated results possible with the use of Nano Banana 2 and be gifted with high-fidelity visuals.
Bottom line
To fix Gemini’s hallucinations, start with prompts 1 and 6. Use 2 for logic, 9 for memory, and 3 to loosen over-active safety filters.
If Gemini gets too chatty, prompts 4 and 7 will keep it concise. To see the model’s full potential—including varied perspectives and its best hidden features — round out your workflow with 5, 8 and 10.
Let me know in the comments if you give any of these a try. I'd love to know what you think.
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom's Guide on YouTube and follow us on TikTok.
More from Tom’s Guide
- I fixed Claude’s biggest flaws — these 10 prompts help improve its answers fast
- I fixed ChatGPT’s biggest weaknesses — these 10 prompts instantly improve its answers
- I asked ChatGPT, Gemini, and Claude how to stop feeling overwhelmed and these small changes actually helped
