How to Write Prompts: Tips to Dramatically Improve Answer Quality

Introduction: Even the Same AI Answers Very Differently by How You Ask

When answers feel "shallow," "off-intent," or "too long," the cause is usually not model performance but prompt (instruction) design. Beginners should first learn three things: role definition, providing context, and specifying the output format. Just these greatly improve accuracy and reproducibility.

AI is good at long-text understanding, summarization, writing, and analysis, but responds vaguely to vague requests. Use it with the mindset of giving a good instruction sheet rather than dumping work on it.

The Basic Structure of a "Good Prompt"

In real work, this template tends to be stable.

You are [role].
Objective: [what you want]
Context: [background, target reader, constraints]
Input: [materials or situation]
Output format: [bullets, table, headings, char count, etc.]
Notes: [what to avoid, what to prioritize]

For an internal email draft, for example:

You are an assistant well-versed in Japanese business documents.
Objective: write an email to coordinate a delivery date with a client.
Context: respectful to the recipient while honestly conveying our situation.
Input: delivery will be 3 days late due to a parts-procurement delay; partial advance delivery is possible.
Output format: three subject options, one body in polite Japanese.
Notes: don't sound like an excuse; include concrete countermeasures.

Just this structure produces far more practical results than "write a delivery-delay email."

The Effect of Improvement, Before / After

Example 1: Blog outline

Before

Come up with an SEO article outline.

Theme, reader, target keywords, and depth are unclear.

After

You are an SEO content editor.
Objective: create a blog outline on "can't focus working from home."
Context: readers are office workers 1–3 years in who work remotely. Search intent: "improvement tips I can try right away."
Output format:
- 3 title ideas
- h2/h3 structure
- 1–2 sentences of key points per heading
Notes: focus on action-level measures, not motivational talk.

The improvement: stating reader, search intent, output format. You get an outline that's actually easy to turn into an article.

Example 2: Summarization

Before

Summarize this document.

After

Summarize the following meeting materials.
Objective: so my manager grasps the overview in 2 minutes.
Output format:
1. Overall summary (within 150 chars)
2. Three key points
3. Items needing a decision
4. Concerns
Paraphrase jargon where possible.

A good summary depends on who reads it. Adding reader and purpose makes it usable.

Role Definition Is Not "Playing Expert" but a Technique to Fix the Viewpoint

"You are a marketer" isn't vibe-setting; it tells the AI which perspective to prioritize.

• Marketer: messaging, target, differentiation
• Editor: readability, structure, avoiding redundancy
• Engineer: accuracy, assumptions, edge cases
• HR: evaluation criteria, fairness of expression, practical operation

The more specific the role, the better. "A B2B marketer at a SaaS company" beats "a marketer."

Reasoning: "Have It Output a Structured Process" Rather Than "Show It"

Rather than dumping a long thought process, having it output a structured decision process is more stable. If you want Chain-of-Thought, write it like:

Not just the conclusion—organize the points needed to judge, in order.
Output in this format:
- Assumptions
- Comparison of options
- Recommended option
- Reasoning

For tool selection:

You are an information-systems staffer.
Objective: compare internal chat tools for SMBs.
Conditions: prioritize cost, manageability, security, Japanese support.
Compare 3 options, then recommend 1.
Output order: "comparison table" → "reason for recommendation" → "cautions on adoption."

Few-shot Is a "Show Examples" Technique

When intent doesn't get through, showing 1–3 examples can be faster than more explanation.

Write the reply in the same tone as the examples.

Example 1
Customer: wants shipping status
Reply: Thank you for contacting us. Your order shipped today and usually arrives in 1–2 days.

Example 2
Customer: wants to return an item
Reply: Thank you for your inquiry. Understood. Returns are possible within 7 days of receipt.

This time: confirm the invoice issuance date

Few-shot helps align tone, classification criteria, output pattern. Too many examples become noise—about 2 is enough.

Step-by-step Breaks Down Complex Work to Reduce Mistakes

A big one-shot request tends to yield a convincing-but-rough answer. Split into stages.

Proceed in this order.
Step 1: organize the problem into 3 items
Step 2: present cause hypotheses for each
Step 3: propose easy-to-execute countermeasures, prioritized
Step 4: conclude with a 1-week action plan

Especially effective for business improvement, learning plans, and planning. Problem framing → cause hypotheses → countermeasures feels more convincing than demanding a methodology immediately.

Comparison of Practical Techniques

A General Template When Unsure

You are [role]. Handle the following.

[Objective] [the final result you want]
[Background] [situation, audience, use]
[Input] [materials, data, notes]
[Output format] [bullets / table / headings / char count]
[Prioritize] [accuracy, conciseness, practicality, etc.]
[Note] If info is missing, ask up to 3 clarifying questions before starting.

The last line is very effective: it prevents charging ahead on guesses and reduces rework.

Summary: A Good Prompt Is an Instruction Sheet for AI

Improving answer quality isn't about complicated spells—it's making clear as whom, for what, in what form you want the output. Start with the three (role, context, output format), then add Few-shot or Step-by-step as needed. It answers even careless asks, but the more carefully you ask, the higher the accuracy. From the next chapter we move to use-case templates and safe usage.