Introduction: Control the Shape of Qwen3’s Responses

When building AI agents, tools, or apps with Qwen3, output format matters.

This guide will show you how to:

• Generate valid JSON for parsing or APIs

• Create tables for reports and analysis

• Use markdown formatting for readable outputs

• Control output precision and consistency

With the right prompts, Qwen3 can act like a reliable backend service—not just a chatbot.

---

1. Basic Prompt Structure

Use OpenAI-style chat format:

text
<|im_start|>system
You are a JSON generator bot. Always return clean JSON.
<|im_end|>
<|im_start|>user
Convert this info into JSON: Name: Alice, Age: 24, Skills: Python, ML
<|im_end|>
<|im_start|>assistant
{
  "name": "Alice",
  "age": 24,
  "skills": ["Python", "ML"]
}

✅ Use system to lock the output format style.

---

2. Generate JSON for APIs & Functions

Prompt Template:

text
<|im_start|>system
Respond only in raw JSON. No extra text or explanations.
<|im_end|>
<|im_start|>user
Create a product object with name, price, stock, and tags.
<|im_end|>

Qwen3 Output:

json
{
  "name": "Smart Watch",
  "price": 199.99,
  "stock": 42,
  "tags": ["electronics", "wearable", "fitness"]
}

---

3. Generate Markdown (Great for Docs, UIs)

Prompt:

text
<|im_start|>system
Format your answer in markdown.
<|im_end|>
<|im_start|>user
Write a comparison of Qwen3 and GPT-4.
<|im_end|>

Output:

markdown
| Feature        | Qwen3          | GPT-4          |
|----------------|----------------|----------------|
| Open Source    | ✅ Yes         | ❌ No          |
| Max Context    | 128K tokens     | 128K tokens    |
| Cost           | Free (local)    | Paid           |

---

4. Generate Tables

Great for BI dashboards, reports, and knowledge summaries.

Prompt:

text
<|im_start|>system
Return a markdown table comparing LLMs by model size and benchmark scores.
<|im_end|>

Output:

markdown
| Model         | Params | HumanEval | MMLU |
|---------------|--------|-----------|------|
| Qwen3-7B      | 7B     | 78.4%     | 62.1 |
| GPT-4-Turbo   | ?      | 90.2%     | 86.0 |
| Claude Sonnet | ?      | 82.3%     | 78.5 |

---

5. Create Structured Instruction Outputs

To guide downstream tools or logic.

Prompt:

text
<|im_start|>system
Return output in this structure: { "task": "", "steps": [], "tools": [] }
<|im_end|>
<|im_start|>user
I want to automate sending emails with Python.
<|im_end|>

Output:

json
{
  "task": "Send emails with Python",
  "steps": [
    "Install smtplib",
    "Compose email content",
    "Connect to SMTP server",
    "Send the email"
  ],
  "tools": ["Python", "smtplib"]
}

---

6. Prompt Engineering Tips

---

7. Common Use Cases for Structured Prompts

---

Conclusion: Structure Is Power

Qwen3 is highly capable at returning clean, structured, and parsable output—you just need to give it the right prompt shape.

Use this for:

• API backends

• Agent chains

• AI dashboards

• Developer tools

With structured prompting, Qwen3 becomes a reliable machine partner, not just a creative writer.

---

Resources

• Qwen Prompt Format

• OpenAI Function Calling Format

• AgentBench (JSON Tools)

• Prompt Engineering Cheatsheet