Tips on four major engineering skills
Basic · Few-shot · Structured · Structured output

1Tip Engineering Overview: Four Progressive Techniques

Prompt Engineering is the most direct means to affect LLM output quality. This section introduces four progressive tips and techniques from the most basic to the most rigorous:

2Basic Tip: Use system_prompt to define Agent roles

The simplest way to prompt: increate_agentwhen passed insystem_promptParameters, tell the model who it is and what it should do.

Default behavior without system_prompt

Without the system prompt, the model will give the "honest" answer - the moon has no capital:

agent = create_agent(model=deepseek_model())

question = HumanMessage(content="What's the capital of the moon?")
response = agent.invoke({"messages": [question]})
print(response['messages'][1].content)

Add system_prompt: switch to science fiction writer role

system_prompt = "You are a science fiction writer, create a capital city at the users request."

scifi_agent = create_agent(
    model=deepseek_model(temperature=1),  # temperature=1 增加创意性
    system_prompt=system_prompt
)

response = scifi_agent.invoke({"messages": [question]})
print(response['messages'][1].content)

3Few-shot tip: embed examples in System Prompt

Few-shot prompting is a powerful technique: give a few "input → output" examples directly in the system prompt, and LLM will imitate the style and format of these examples to generate new responses.

Add Few-shot example

system_prompt = """

You are a science fiction writer, create a space capital city at the users request.

User: What is the capital of mars?
Scifi Writer: Marsialis

User: What is the capital of Venus?
Scifi Writer: Venusovia

"""

scifi_agent = create_agent(
    model=deepseek_model(),
    system_prompt=system_prompt
)

question = HumanMessage(content="What is the capital of The Moon?")
response = scifi_agent.invoke({"messages": [question]})
print(response['messages'][1].content)

Note the change in output style: the model learned the format "Scifi Writer: [word]" in the example and gave a concise one-word answer instead of a long description.

4Structural tip: Use format templates to constrain output structure

Structured Prompting defines clearly field names and output formats in the system prompt, making the model's natural language output more regular and easier to parse.

Define output format template

system_prompt = """

You are a science fiction writer, create a space capital city at the users request.

Please keep to the below structure.

Name: The name of the capital city

Location: Where it is based

Vibe: 2-3 words to describe its vibe

Economy: Main industries

"""

scifi_agent = create_agent(
    model=deepseek_model(),
    system_prompt=system_prompt
)

response = scifi_agent.invoke({"messages": [question]})
print(response['messages'][1].content)

5Structured output: forcing output of JSON objects with Pydantic

Structured Output is the most reliable way: usePydanticDefine the data model byresponse_formatThe parameter tells LLM that the output must be according to this schema. LangChain will automatically parse the model's JSON output into Pydantic objects under the hood.

Define the Pydantic data model

from pydantic import BaseModel

class CapitalInfo(BaseModel):
    name:     str  # 城市名称
    location: str  # 地理位置
    vibe:     str  # 氛围描述
    economy:  str  # 经济支柱

Pass in the response_format parameter

agent = create_agent(
    model=deepseek_model(),
    system_prompt="You are a science fiction writer, create a capital city at the users request.",
    response_format=CapitalInfo   # 关键：传入 Pydantic 类
)

question = HumanMessage(content="What is the capital of The Moon?")
response = agent.invoke({"messages": [question]})

# structured_response 是 Pydantic 对象，不是字符串
print(response["structured_response"])

underlying working mechanism

when you pass inresponse_format=CapitalInfoWhen, LangChain will:

1. Convert the Pydantic model to JSON Schema and inject it into the tool definition
2. LLM received instructions: it must call theCapitalInfoTools, parameters strictly conform to the schema
3. The model outputs tool_call (instead of ordinary text), and the parameters are structured JSON
4. LangChain parses JSON into Pydantic objects and storesresponse["structured_response"]

6Reading and using structured_response

Once you have the Pydantic object, you can read each field just like accessing the properties of a normal Python object:

capital_info = response["structured_response"]

# 访问具体字段
print(capital_info.name)      # → 'Lunar Prime'
print(capital_info.location)  # → 'Mare Imbrium, near the lunar equator'
print(capital_info.vibe)      # → 'futuristic, serene, industrious'
print(capital_info.economy)   # → 'Helium-3 mining...'

# 组合使用
print(f"{capital_info.name} is located at {capital_info.location}")

messages internal structure (advanced debugging)

useresponse_formathour,messagesExcept in the listHumanMessageandAIMessage, there will be one moreToolMessage：

for msg in response["messages"]:
    pprint(msg)

7Comparison and selection guide of four techniques