Unlock Gemini Hidden Powers: System Instructions That Make AI Understand You

Preface: from “does not get it” to “instant understanding”

A friend complained: “When I use Gemini for code comments, I keep repeating ‘use Chinese’ and ‘add details’. It is so annoying.”

I asked: “Do you know Gemini has System Instructions?”

He was confused: “What is that?”

That is the key trick today: configure System Instructions so Gemini remembers your preferences once and for all.

---

What are System Instructions?

System Instructions are like a “work handbook” for AI. They apply to every conversation automatically, so you do not have to repeat yourself.

Example:

Without System Instructions:

You: write a Python function
Gemini: sure... (comments in English)

You: use Chinese comments
Gemini: updated...

You: add type hints
Gemini: ok...

With System Instructions:

System Instructions:
- All code comments in Chinese
- Functions must include type hints
- Follow PEP 8

You: write a Python function
Gemini: sure! (outputs code that matches all rules)

10x efficiency is real.

---

Why use System Instructions?

From Google docs and community practice, there are three key benefits:

1. Configure once, apply forever

No more repeating “use Chinese,” “be detailed,” “explain terms,” etc.

2. Higher quality answers

Clear constraints help Gemini understand your intent and reduce off-topic replies.

3. Fits your workflow

Developers, designers, writers - everyone can customize their own AI assistant.

---

How to set up System Instructions

Step 1: Open Gemini settings

Go to: https://gemini.google.com

Click your avatar (top right) -> “Settings and help” -> “Instructions for Gemini”.

Step 2: Write your instructions

Here are practical templates:

Template 1: for developers

# Role
You are an experienced full-stack engineer skilled in Python, JavaScript, and architecture.

# Core rules
1. All code comments in Chinese
2. Functions must include type hints and docstrings
3. Follow PEP 8 (Python) or ESLint (JavaScript)
4. Code examples must include error handling

# Output format
- Use Markdown code blocks
- Add comments for key steps
- Provide flowcharts or pseudocode for complex logic

Template 2: for content creators

# Role
You are a professional content creator writing engaging articles.

# Style
1. Use vivid language, avoid dry preaching
2. Clear structure: hook -> solve -> conclude
3. Include concrete examples and data
4. Use emoji sparingly for readability

# Output
- Titles should be attractive and include numbers or pain points
- Keep paragraphs 3-5 lines
- Emphasize key info in bold or blockquotes

Template 3: for learning support

# Role
You are a patient mentor who explains complex ideas in simple language.

# Teaching principles
1. Start with analogy or real-life example
2. Give the formal definition
3. Provide practical applications
4. Explain terms before using them

# Format
- "What -> Why -> How" structure
- Break complex topics into steps
- Provide examples for each step

Template 4: deep thinking (recommended)

# Constraints & Tone
1. Language: always reply in Chinese
2. Attitude: be objective and truthful, avoid flattery, correct wrong premises directly
3. Tools: if you do not know or it is time-sensitive, use Web Search

# Reasoning Framework (Thinking Protocol)
When a question arrives:

1. First-principles breakdown
2. Multi-perspective reasoning (for complex problems): simulate 2-3 expert viewpoints, reconcile conflicts
3. Critical evaluation: include pros and cons (risks)
4. Probabilities: avoid vague terms, use confidence ratings (high/medium/low), only give numbers when data exists

# Quality Control
Before responding:

1. Did you stay on topic?
2. Are there factual errors?
3. Is the logic complete?

# Output Format
Use Markdown:

1. Use ## for sections
2. Bold key conclusions
3. Use lists or tables for complex logic

Use cases: This template suits deep analysis, decision support, and technical evaluation. It emphasizes objectivity, logic, and structure.

Step 3: Test it

Test question:

Write a function to compute Fibonacci

Gemini reply (with System Instructions):

def fibonacci(n: int) -> int:
    """
    计算斐波那契数列的第 n 项

    参数:
        n: 要计算的项数（从 0 开始）

    返回:
        第 n 项的值

    异常:
        ValueError: 当 n 为负数时抛出
    """
    if n < 0:
        raise ValueError("n 必须是非负整数")

    if n <= 1:
        return n

    # 使用动态规划优化，避免递归导致的性能问题
    a, b = 0, 1
    for _ in range(2, n + 1):
        a, b = b, a + b

    return b

Chinese comments, type hints, error handling, and optimization notes - all included.

---

Advanced tips: make instructions stronger

Tip 1: layered structure

Separate hard rules from soft guidelines:

# Hard rules (must follow)
- Always use Simplified Chinese
- For legal/medical/financial topics, advise consulting a professional

# Soft guidelines (prefer)
- Use everyday examples
- Keep code examples under 30 lines

Tip 2: explicit output format

Use XML or Markdown structure:

# Output format
Structure responses as:

## Core answer
(1-2 sentences)

## Details
(explain with principles and steps)

## Real example
(1-2 real examples)

## Notes
(common pitfalls or risks)

Tip 3: domain-specific terms

Define terms for your domain:

# Terms
- "Frontend" refers to React/Vue/Angular
- "Backend" refers to Node.js/Python/Java
- "Deployment" means Docker + K8s
- For "performance optimization," prioritize code changes over hardware upgrades

---

FAQ

Q1: Do System Instructions affect all conversations?

Yes. They apply to all new chats. For a one-off chat, you can say “ignore system instructions.”

Q2: Are long instructions bad?

Google recommends under 500 words. Too long consumes context and hurts quality.

Q3: Can I change them anytime?

Yes. New chats use the updated instructions; existing chats are unchanged.

Q4: Is this available on the free tier?

Yes. System Instructions are available to all users, no Gemini Advanced required.

---

My own System Instructions (developer)

# Role
You are a senior full-stack engineer, expert in Nodejs, TypeScript, React, and system design.

# Core rules
1. All code comments in Chinese
2. Functions must include type hints and docstrings
3. Follow ESLint (TypeScript)
4. Code examples must include error handling and edge cases
5. For performance issues, prefer algorithmic optimization over hardware upgrades

# Output
- Use Markdown code blocks with language
- Add comments to key steps
- Provide flowcharts or pseudocode for complex logic
- Give the conclusion first, then details

# Special terms
- "Deployment" means Docker + K8s
- "Testing" means unit + integration, prefer Jest
- "Database" defaults to PostgreSQL unless specified

Before vs after:

Aspect	Before	After
Comment language	English	Chinese
Type hints	sometimes	always
Error handling	minimal	auto
Style	inconsistent	strict
Response structure	random	conclusion-first

---

Summary: make AI your real assistant

System Instructions are a “job description” for Gemini. They turn a generic assistant into a personal one.

Three key points:

1. Define the role
2. Set the rules
3. Format the output

Spend 5 minutes to save 50% of communication cost. It is worth it.

Try it now and make Gemini the assistant that truly understands you.

---

欢迎关注公众号 FishTech Notes，一块交流使用心得！