Skip to content

Prompt Management

This chapter covers writing and managing the prompts that drive an agent's response quality.

Two Kinds of Prompts

Type Korean Role
System Prompt System Prompt Defines the AI's role, tone, and response format. Hidden from end users
User Prompt User Prompt The question or request from the user (or a template thereof)

System prompts shape an agent's "personality"; user prompts are the input for each conversation.

Prompt Library

Select Agent Creation → Agent Prompt in the left sidebar.

Prompt Library — Storage / Shared / Store tabs with prompt cards, search, and create button

Tabs:

Tab Description
Storage Prompts you created
Shared Prompts others shared with you
Store Team/org-wide shared library

Creating a New Prompt

  1. Click + New Prompt at the top right
  2. Enter:
    • Name
    • Type: System Prompt / User Prompt
    • Content: actual prompt text
    • Variables (optional): placeholders in the form {{ name }} for dynamic insertion
    • Tags (optional): for classification
  3. Save

Prompt editor — the editing view opened by *New Prompt* (name / type / content / variables / tags)

Button label

The actual solution button label is "새 프롬프트" / "New Prompt" (without the + prefix that earlier versions of this manual mentioned).

Example with Variables

You are an expert in {{ role }}.
Answer {{ question }} in a {{ style }} tone.

{{ role }}, {{ question }}, and {{ style }} are replaced with actual values at use time.

Uploading to the Store

To share your prompt with the team:

  1. Prompt detail → Upload to Store
  2. Choose visibility (Company / Department / Specific group)
  3. Upload

Using a Template

Quickly start from a vetted prompt in the library:

  1. Search in the Store tab
  2. Click Duplicate on the prompt card → copies into your storage
  3. Edit as needed

Writing Tips

  • Write clearly and concretely — Instead of abstract directions like "Answer well", pair them with concrete conditions such as "Explain in three sentences or fewer, in a way that's easy for a user to understand."
  • Specify the output format — Naming the desired response shape (JSON, table, numbered list, etc.) yields more consistent results.
  • State the constraints, too — To reduce speculative answers or unnecessary output, pair the request with prohibited conditions. Example: "Do not guess; if you're unsure, reply that the answer needs to be verified."
  • Include examples — Providing one or two examples close to the desired answer shape improves response quality and consistency.

Operational Recommendations

  • Version control — Duplicate before major changes as a backup
  • A/B testing — Run two versions on the same input and compare
  • No sensitive content — Do not embed passwords, API keys, or customer information directly in prompts; isolate via variables

Contact

For prompt-related questions, please contact the Xgen Solution Administrator.