Prompt Basics: How to Ask AI and Get Good Answers

“Prompt engineering” sounds fancy, but the core job is simple: turn vague intent into instructions the model can execute.

Technique 1: Role and Background

❌ Write me an email
✅ You are a B2B sales rep with 10 years of experience. Write a follow-up email
   to a CTO who attended our demo last week but hasn't replied.
   Professional tone, no pressure, under 150 words.

Role setting is not mysticism—it activates domain-appropriate patterns from training data.

Technique 2: Specify Output Format

❌ Analyze this competitor
✅ Analyze the competitor below as a table with columns:
   Features, Pricing, Strengths, Weaknesses, Target Users

   | Features | Pricing | Strengths | Weaknesses | Target Users |
   |----------|---------|-----------|------------|--------------|
   | ...      | ...     | ...       | ...        | ...          |

Models excel at structured output. Specify format and you avoid wall-of-text prose.

Technique 3: Examples (Few-Shot)

Often better to show one example than describe at length:

Classify user feedback as: Feature Request / Bug Report / Usage Question

Examples:
Input: "Can you add dark mode?" → Feature Request
Input: "Login page keeps spinning" → Bug Report

Now classify:
Input: "Where is your API documentation?"

One example often beats three paragraphs of description.

Technique 4: Step by Step (Chain of Thought)

For complex tasks, ask the model to proceed in steps:

❌ Should I change jobs?
✅ Analyze my job-change decision in steps:
   1. List pros/cons of current role (3 each)
   2. List pros/cons of new offer (3 each)
   3. Compare on career growth, income, work-life balance
   4. Give analysis and recommendation (mark inferences clearly)

Benefits: structured output and visible reasoning (easier to spot logic gaps).

Technique 5: System Prompt as Standing Rules

When building AI apps, the system prompt is the model’s job description:

You are a code review assistant. Rules:
- Review only security and performance; ignore style
- Every issue must include line number and fix suggestion
- If no issues, reply "LGTM"
- Do not invent problems that don't exist

System prompts persist across the conversation—more efficient than repeating rules every turn.

Common Pitfalls

Pitfall	Cause	Fix
Generic answers	Question too open	Add constraints: length, format, audience
Contradictions	Context truncated	Trim history or summarize
Fabricated facts	Prediction, not lookup	Provide source material or RAG
Unstable output	Temperature too high	Lower temperature; add format constraints

A Complete Prompt Template

[Role] You are…
[Background] The situation is…
[Task] Please…
[Format] Output as…
[Constraints] Do not…; must…
[Example] For instance…

You need not fill every slot—but when nothing works, check what’s missing.

Summary

Prompts are not spells—they are clear requirements. Role + format + examples + steps cover ~80% of daily use.

Next: how to make AI read your documents before answering—RAG.