What is Prompt Engineering? The Art of Designing How to Communicate with AI
Prompt engineering is instruction design to reliably produce the desired outputs from generative AI (such as ChatGPT). It may sound difficult, but it’s basically like writing a well-structured brief when you ask a person to do a task.
AI may seem万能, but ambiguous requests are what it struggles with most. Conversely, when the objective, assumptions, constraints, and evaluation criteria are all in place, it moves remarkably intelligently. This article summarizes templates and techniques you can apply starting tomorrow.
The Basic Structure You Should Lock In First: The Five-Point Set of a Good Prompt
Prompts matter more in terms of how the information is assembled than length. If you’re unsure, try including these five points.
- Goal: What is this for? (e.g., improving CVR, writing specifications, learning)
- Role: Who should the AI be for? (e.g., editor, SRE, recruiter)
- Input: What are the materials? (e.g., meeting minutes, logs, text, requirements)
- Constraints: Conditions to follow (e.g., length, tone, prohibitions)
- Output format: How would you like the response? (e.g., table, bullet list, JSON)
Template (Copy-paste OK)
Your role: Act as an expert in 〇〇.
Goal: We want to achieve 〇〇.
Background/Assumptions: 〇〇. The intended audience/users are 〇〇.
Input: Use the following information: …
Constraints: ・… (prohibited/mandatory/character limit/tone)
Output format: Use headings → bullet list → conclusion order. If a table is preferred, provide a Markdown table.
Prompt Techniques That Work Well (In Practical Order of Effectiveness)
1) Include a concrete example (Few-shot)
AI tends to be more stable with examples than with abstract instructions. For instance, by providing one completed example, you can align style and format.
Example: Please create three in this format.
【Example】Task: ~ / Cause: ~ / Countermeasure: ~ / Expected effect: ~
2) Start by asking to 'Ask first, then proceed'
When requirements are unclear, it’s wise not to let the AI run ahead. Having it ask clarifying questions first reduces back-and-forth.
