What Is Prompt Engineering? The Skill of Designing "How to Say It So AI Understands"
Prompt engineering is instruction design for getting generative AI (such as ChatGPT) to stably produce the output you want. It sounds difficult, but it is essentially close to "writing a good request document when you ask a person to do a job."
AI looks all-powerful, but it is actually worst at vague requests. Conversely, when purpose, premises, constraints, and evaluation criteria are all present, it works surprisingly smartly. In this article, we collect templates and techniques you can use starting tomorrow.
The Basic Structure to Master First: The "Set of 5" for a Good Prompt
For prompts, how well the information is assembled matters more than length. If you are unsure, try including the following set of 5.
- Purpose: What is it for? (e.g., improve CVR, write a spec, learning)
- Role: Who do you want the AI to be? (e.g., editor, SRE, recruiter)
- Input: What is the material? (e.g., meeting minutes, logs, text, requirements)
- Constraints: Conditions to obey (e.g., character count, tone, prohibitions)
- Output format: How do you want it returned? (e.g., table, bullets, JSON)
Template (OK to copy-paste)
Your role: Please act as an expert in XX.
Purpose: I want to achieve XX.
Background / premises: It is XX. The assumed reader/user is XX.
Input: Please use the following information: ...
Constraints: ... (prohibited / required / character count / tone)
Output format: In the order heading then bullets then conclusion. Use a Markdown table when a table is desirable.
A Collection of Prompt Techniques That Work (In Order of How Much They Help in Practice)
1) Include one "concrete example" (Few-shot)
AI suddenly stabilizes when you show an example rather than an abstract instruction. For instance, writing style and format align just by presenting one finished form.
Example: Please create 3 in this format.
[Example] Issue: ~ / Cause: ~ / Countermeasure: ~ / Expected effect: ~
2) First ask it to "ask questions before working"
When requirements are vague, the trick is not to let the AI run off. Having it ask clarifying questions first reduces rework.
If there is missing information, please ask up to 5 questions first. After my answers, please produce the final draft.
3) Make the evaluation criteria explicit (have it self-grade)
Instead of "make it nice," write what counts as good. AI optimizes more easily when there are evaluation criteria.
Evaluation criteria: (1) even a beginner won't misunderstand (2) conclusion first (3) two or more concrete examples (4) technical terms supplemented in a few words
4) Write constraints separating "prohibited" and "required"
The more constraints there are, the better they are obeyed if you separate prohibited items and required items in bullets.
- Required: conclusion at the top, steps numbered, next action at the end
- Prohibited: definitive legal advice, numbers of unknown origin, fabricating company names
5) Have it "structure the output" (JSON / table / headings)
If you use it for work, a form that is easy to pass downstream is strongest. In particular, JSON flows easily into other tools (Notion / spreadsheets / your own system) and improves reusability.
Output in JSON. keys are title, problem, solution, risks, nextActions. In Japanese.
By Use Case: Prompt Examples You Can Use As-Is
Writing (blog / internal documents)
You are an editor. The purpose is to write "an article even a beginner can understand."
Theme: the basics of prompt engineering
Assumed reader: non-engineer business roles
Required: supplement technical terms with a word in parentheses. Use headings equivalent to h2/h3. Three concrete examples.
Output: structure proposal then body then 5 key points at the end.
Summarization (minutes / papers / long text)
Please summarize the following.
Constraints: separate (1) important decisions (2) open items (3) owners and deadlines. Each within 3 lines.
At the end, up to 5 "questions to check before next time."
Ideation (planning / initiatives)
You are a PM for B2B SaaS. The purpose is onboarding improvement.
Premise: 14-day free trial, current activation rate is 22%.
Output: 10 initiatives. For each, attach "aim," "implementation difficulty (low/med/high)," and "measurement metric (e.g., Day1 active rate)."
Coding support (design review / debugging)
For code, environment, constraints, and reproduction steps are key. Paste the error log too if possible.
You are a senior engineer. Please tell me the cause and a fix for the following error.
Environment: Python 3.11 / FastAPI / Docker
Reproduction steps: ...
Expected: /health returns 200
Actual: 500, log is as follows: ...
Output: candidate causes in priority order, verification commands, an example fix patch.
Patterns That Tend to Fail, and Easy Fixes
- Vague: "make it nice" then add who it's for, purpose, output format
- Insufficient information: no background then add premises, constraints, usage scenario
- Aiming for perfect in one shot: a long request then split into 3 stages: structure then draft then refine
- Not verifying: adopting the output then always ask for basis, counterexamples, risks
Prompts Are "Grown Through Conversation": The Iteration Template (3 Steps)
- Draft: first get a 60-point answer out (prioritize speed)
- Review: pass it viewpoints and have it self-check (e.g., gaps/contradictions/risks)
- Improve: fix with diff instructions (accuracy improves more easily than a full rewrite)
Please review this output from the viewpoints of (1) reader misunderstanding (2) logical leaps (3) lack of specificity, and produce a revision. Make the revision in a form where the diff is visible.
Latest Trends: Not Just Prompts—"Surrounding Design" Has Become Important
Recently, beyond just "trying hard on prompts," there is a strong flow of stabilizing quality on the system side. The representative three are the following.
- RAG: search internal documents and pass the basis (a measure against hallucination)
- Tool calling (Function Calling): let the AI do calculation, search, and DB lookups to raise accuracy
- Evaluation (Evals): create tests of good/bad output and continuously improve
In other words, prompts are not "the last magic"; they are strong when treated as part of reproducible operations.
Summary: A Shortcut to Improving Prompts That Works From Tomorrow
People who are good at prompts are good at request design rather than having a sense for phrasing. First, assemble the set of 5 (purpose, role, input, constraints, output format), show an example, have it ask questions, and bind it with evaluation criteria. With just this, accuracy improves visibly.
Finally, if you tell me your use case (writing, summarizing, planning, code), we can build a dedicated template together.
