5 ‘Prompt Cheat Codes’ I Use When Gemini Misbehaves
Recently, trying to increase work efficiency with Gemini often leads to more stress instead. I clearly gave instructions in Korean, but the results are strangely awkward or completely off-topic. If this repeats, I end up thinking, “I’ll just do it myself.”
So I changed my mindset. Instead of blaming the AI’s performance, I decided to change the way I ask questions. In this post, I am sharing my “5 Fail-Safe Prompt Formulas” that I use whenever AI responses aren’t up to par.
Gemini Prompt Guide: Boosting Answer Quality
1. One “Answer Key” is Better than a Hundred Explanations (One-shot)
“Friendly but a bit cynical, yet with a warm aftertaste.” Humans can’t understand abstract expressions like this. AI understands them even less.
What I wanted wasn’t a long explanation, but a result that felt “exactly like this.” The most certain way is to throw it one well-written example (Sample).
[Bad] Listing ambiguous adjectives
- Instruction: “Write a funny but slightly poignant Instagram post in a trendy tone.”
[Good] Presenting a model answer (One-shot)
- Instruction: (Paste a post I wrote that had good reactions) “The text above is my usual style. Keeping this tone and nuance exactly, write a new post about ‘commuting’.”
2. Don’t Just Say ‘What’, Say ‘Why’ (Goal)
If you ask AI for ideas, it just gives you clichés. The reason is simple.
I didn’t tell it the final goal. You need to inform it of the result you want to achieve through this output.
[Bad] Simple request without a purpose
- Instruction: “Give me 10 ideas for Instagram posts.”
[Good] Setting a specific Goal
- Instruction: “Think of 10 Instagram topics for our company. The goal this time isn’t just ‘likes’, but to get people to click the profile link and come to the product page. Focus on curiosity-inducing ideas.”
3. Don’t Ask for Everything at Once, Ask for ‘One Thing’ (Next Action)
As the importance of prompts becomes known, various prompts are being shared.
I also create and use complex prompts, but if I ask for too many things at once, the quality sometimes drops.
Show the big picture, but the key is to direct only the “Next Action needed right now.”
[Bad] Demanding the entire process at once
- Instruction: “Plan a marketing campaign for the new product, and organize the schedule and budget all at once.”
[Good] Step-by-step approach
- Instruction: “We’re doing a new product SNS campaign. Target is 20-30s, budget is $5,000. Okay, what should we do first? Let’s start by defining the core message we need to finalize most urgently.”
4. Fix it? No, Scrap it and Rewrite (Rewrite)
When the answer is odd, we often prescribe patch-ups like “Add this to that,” or “No, make it more polite.” As you’ll know if you try, this makes the result even more twisted. This is because conditions conflict.
If a correction is needed, don’t give additional instructions; it’s much faster to organize your intent from the beginning and instruct it to Rewrite.
[Bad] Endless chain of corrections
- Initial Instruction: “Write a team dinner email. Make it comfortable so it doesn’t feel stuffy.”
- Additional Instruction: “Ah, the manager is coming too, so it can’t be too light. Change it to formal language.”
- Result: A strange hybrid comes out, like “Being late is strictly forbidden, ya know?”
[Good] Instruction with integrated intent
- Revised Instruction: “Since the manager is attending, rewrite the welcome dinner announcement email in a tone that maintains formalities but is not too stiff, and is appropriately casual.”
- Result: An email with the right mix of politeness and wit comes out in one go.
5. Find and Fix it Yourself (Self-Correction)
AI is good at lying shamelessly. So you shouldn’t just trust it blindly. Instead, ask the AI back, “What is lacking in your answer?”
Making it censor itself significantly boosts quality.
[Good] Self-criticism prompt
- Instruction: “Find the Top 3 areas that can be further improved in the strategy you just proposed, and suggest improvements together.”
- Result: It upgrades its own answer, saying things like “Numerical evidence is lacking.”
Conclusion
Just doing this produces fairly stable answer quality.
- Show a good sample (One-shot)
- Tell it why we are doing this (Goal)
- Ask for things one by one (Next Action)
- If it’s weird, say it again from the start (Rewrite)
- Make it review itself (Self-Correction)
If your work is blocked because of AI answers, it might be worth applying these 5 methods at least once.
