Prompt Architecture & Reusability

1. The "Magic String" Problem

In a demo, your prompt is 3 lines. In production, your prompt is 3 pages.

If you keep 3-page strings inside your Python files (main.py), your code becomes unreadable. You can't version control the logic separately from the text.

2. Solution: The Prompt Registry Pattern

Treat prompts like assets. Create a folder structure:

src/
  prompts/
    __init__.py
    support_agent.py
    sales_agent.py

Inside support_agent.py:

from langchain_core.prompts import ChatPromptTemplate

SYSTEM_TEMPLATE = """
You are a support agent for company X.
Current Time: {time}
Rules:
1. Be polite.
2. Never promise refunds.
"""

def get_support_prompt():
  return ChatPromptTemplate.from_messages([
      ("system", SYSTEM_TEMPLATE),
      ("human", "{input}")
  ])

This seems simple, but it allows you to:

1. Import the prompt anywhere: from prompts.support_agent import get_support_prompt
2. Test the prompt in isolation.
3. Diff changes easily in Git (text files).

3. Composition (Layered Prompts)

Sometimes you have a "Base Prompt" (Company Tone) and a "Specific Prompt" (Task). You can compose them.

BASE_SYSTEM = "Always speak in a professional tone."

TASK_SPECIFIC = "User is asking about billing."

full_prompt = ChatPromptTemplate.from_messages([
  ("system", BASE_SYSTEM),
  ("system", TASK_SPECIFIC), # Yes, you can stack system messages
  ("human", "{input}")
])

4. Summary

Don't bury prompts in your logic. Externalize them. Modularize them.

Key Intuition: "Prompts should evolve like code."