Prompt Engineering for Chinese AI Models

Prompt engineering is the difference between mediocre and exceptional AI outputs. Chinese AI models have unique characteristics that require tailored prompting strategies. This guide covers proven techniques for DeepSeek, GLM, Qwen, Kimi, and Doubao to help you extract the best possible results.

Why Chinese Models Need Different Prompting

Chinese-trained models process language differently from English-centric models:

• Context sensitivity: Chinese relies heavily on context and implication. Vague prompts produce vague answers.
• Honorifics and tone: Chinese has formal (敬语) and casual registers. Prompts should specify the desired tone.
• Idioms and chengyu: Chinese models excel at 成语 and cultural references but need explicit permission to use them.
• Number handling: Chinese numbers (一, 二, 三) vs Arabic numerals (1, 2, 3) — specify which format you need.

Core Techniques

1. Role Prompting (角色设定)

Assigning a specific role dramatically improves output quality for Chinese models:

import requests

API_KEY = "your-tokenease-api-key"

def generate_with_role(model, role, task, constraints=""):
    messages = [
        {
            "role": "system",
            "content": f"你是{role}。请严格按照用户要求完成任务，不要添加额外解释。"
        },
        {
            "role": "user",
            "content": f"{task}\n\n要求：{constraints}" if constraints else task
        }
    ]
    
    response = requests.post(
        "https://tokenease.io/v1/chat/completions",
        headers={"Authorization": f"Bearer {API_KEY}"},
        json={
            "model": model,
            "messages": messages,
            "temperature": 0.7
        }
    )
    return response.json()["choices"][0]["message"]["content"]

# Usage
result = generate_with_role(
    model="zhipu",
    role="资深电商文案策划",
    task="为智能手表写产品文案",
    constraints="三段，每段≤100字，风格：科技/情感/性价比"
)

2. Few-Shot Prompting (示例引导)

Chinese models respond exceptionally well to examples:

messages = [
    {"role": "system", "content": "你是一个文本分类器。请根据示例判断用户意图。"},
    {"role": "user", "content": "示例1：\"明天北京天气怎么样？\"\n意图：查询天气"},
    {"role": "assistant", "content": "查询天气"},
    {"role": "user", "content": "示例2：\"帮我订一张去上海的机票\"\n意图：预订机票"},
    {"role": "assistant", "content": "预订机票"},
    {"role": "user", "content": "示例3：\"这个手机的电池能用多久\"\n意图："}
]

# The model will output: 查询产品信息

3. Chain-of-Thought (思维链)

For complex reasoning tasks, ask the model to think step by step:

prompt = """请逐步分析以下数学问题，展示你的思考过程：

问题：一个水池有两个进水管和一个排水管。甲管单独注满需6小时，乙管单独注满需4小时，排水管单独排空需3小时。如果三个管同时打开，多久能注满水池？

请按以下格式回答：
1. 分析已知条件
2. 建立数学模型
3. 逐步计算
4. 给出最终答案"""

messages = [
    {"role": "user", "content": prompt}
]

# DeepSeek and Kimi excel at CoT reasoning
response = requests.post(
    "https://tokenease.io/v1/chat/completions",
    headers={"Authorization": f"Bearer {API_KEY}"},
    json={
        "model": "deepseek",
        "messages": messages,
        "temperature": 0.3  # Lower temp for reasoning
    }
)

4. Structured Output Prompting

Force structured outputs without relying solely on JSON mode:

prompt = """从以下新闻中提取关键信息，严格按照指定格式输出：

新闻：{news_text}

输出格式：
- 标题：[新闻标题]
- 关键人物：[逗号分隔的人物列表]
- 事件时间：[YYYY-MM-DD格式]
- 事件地点：[地点]
- 核心事实：[一句话概括]
- 影响评估：[1-5分，1=影响很小，5=影响巨大]

注意：只输出上述格式内容，不要添加任何额外说明。"""

# This works better than pure JSON instructions for Chinese models

Model-Specific Prompting Tips

Model	Prompting Style	Temperature
DeepSeek-V4	Explicit step-by-step instructions. Loves reasoning tasks.	0.3-0.5
GLM-5.1	Rich context and role definition. Excels at creative writing.	0.7-0.9
Qwen-Plus	Direct, concise prompts. Fastest response to clear instructions.	0.5-0.7
Kimi-K3	Long context windows. Great for document analysis prompts.	0.3-0.7
Doubao-Pro	Structured, business-oriented prompts. Prefers formal tone.	0.5-0.7

Advanced Patterns

Self-Consistency Voting

Generate multiple answers and pick the most common one for critical tasks:

def self_consistent_answer(model, prompt, n_samples=5):
    answers = []
    for _ in range(n_samples):
        response = requests.post(
            "https://tokenease.io/v1/chat/completions",
            headers={"Authorization": f"Bearer {API_KEY}"},
            json={
                "model": model,
                "messages": [{"role": "user", "content": prompt}],
                "temperature": 0.8  # Higher for diversity
            }
        )
        answers.append(response.json()["choices"][0]["message"]["content"])
    
    # Simple voting: return most common answer
    from collections import Counter
    return Counter(answers).most_common(1)[0][0]

Prompt Chaining

Break complex tasks into sequential prompts:

def chain_prompts(model, user_request):
    # Step 1: Plan
    plan_prompt = f"用户请求：{user_request}\n请制定一个3步执行计划，每步一句话。"
    plan = call_model(model, plan_prompt)
    
    # Step 2: Execute each step
    steps = plan.split('\n')
    results = []
    for step in steps:
        result = call_model(model, f"执行以下步骤：{step}")
        results.append(result)
    
    # Step 3: Synthesize
    synthesis_prompt = f"综合以下结果，给出最终答案：\n{chr(10).join(results)}"
    return call_model(model, synthesis_prompt)

Negative Prompting

Tell the model what NOT to do:

Common Mistakes

Mistake	Why It Fails	Fix
Overly vague prompts	Chinese models need context	Add role, constraints, format
Mixing languages	Confuses tokenization	Stick to one language per prompt
No output format	Inconsistent structure	Specify exact format with examples
Wrong temperature	Creative tasks need high temp; reasoning needs low	Match temp to task type
Ignoring system prompt	Wastes context window	Put instructions in system prompt

Production Prompt Template

Use this template for consistent, high-quality prompts:

SYSTEM_PROMPT = """你是{role}。

任务：{task_description}

约束条件：
- 输出长度：{length_constraint}
- 输出格式：{format_spec}
- 语气风格：{tone}
- 禁止事项：{negative_constraints}

示例输出：
{example_output}

请严格按照以上要求完成任务，不要添加额外内容。"""

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": user_input}
]

Conclusion

Prompt engineering for Chinese AI models follows the same principles as English models but requires attention to linguistic and cultural nuances. Role prompting, few-shot examples, and structured output formats are your most powerful tools.

Start with a clear system prompt defining the role and constraints. Add few-shot examples for complex tasks. Use chain-of-thought for reasoning. Test across models to find the best fit. With practice, you'll consistently get production-quality outputs from Chinese AI models.