Module 02Lesson 2

Lesson 2. The Prompt as a Technical Specification

Theory

Lesson 2. The Prompt as a Technical Specification#

Prompt — instruction for AI
Prompt — instruction for AI

Why This Matters#

A prompt isn't just a question. It's an instruction to the model on how to work. A good prompt = a quality answer.

Key Idea#

Prompt = role + task + data + format

The clearer you formulate, the better the result.

Structure of a Good Prompt#

1. Role

Who is the model in this situation?

Examples:

  • "You are a customer support manager"
  • "You are a marketing expert"
  • "You are an executive assistant"

Why: the role sets the style and tone of the answer.

2. Task

What should the model do?

Examples:

  • "Answer the client's question about delivery"
  • "Analyze reviews and highlight the main issues"
  • "Draft a letter to the client"

Why: the task defines the action.

3. Input Data (Context)

What information is needed for the answer?

Examples:

  • "Delivery terms: 2–5 days, free from $50"
  • "Client reviews: [list]"
  • "Client info: order #123, paid"

Why: data = facts for the answer.

4. Response Format

How should the result look?

Examples:

  • "Answer briefly, 2–3 sentences"
  • "List issues as a bullet list"
  • "Format the letter with greeting and signature"

Why: format makes the answer usable.

Example of a Bad Prompt#

"Answer the client about delivery"

Problems:

  • no context (what are the delivery terms?)
  • no format (how to answer?)
  • no constraints (what to say if we don't know?)

Model's answer: "Hello! Delivery usually takes 1–3 days, free from $30."
← may be incorrect

Example of a Good Prompt#

You are a support manager for an online clothing store.

Task: answer the client's question about delivery.

Delivery terms:
- Nationwide: 3–7 days, free from $50, otherwise $3
- Local: 1–2 days, free from $30, otherwise $2
- Delivery by courier or to pickup point

Response format:
- Polite tone
- 2–4 sentences
- If the client didn't specify city, ask

Client's question: {question}

Model's answer: "Hello! Nationwide delivery takes 3–7 days, free for orders over $50, otherwise $3. For local delivery we deliver in 1–2 days, free from $30. Please specify your city for accurate information."

Prompt Template#

You are [role].

Task: [what to do].

Context/data:
- [information 1]
- [information 2]
- [information 3]

Rules:
- [constraint 1]
- [constraint 2]

Response format:
- [how to format]

Input: [question/text/data]

Example Prompts for Different Tasks#

Example 1: FAQ Bot

You are an assistant for the beauty salon "Beauty Spot".

Task: answer clients' frequent questions.

Knowledge base:
- Hours: Mon–Fri 10:00–20:00, Sat–Sun 11:00–19:00
- Address: 123 Main St
- Services: haircut, coloring, manicure, pedicure
- Prices: haircut from $15, coloring from $30

Rules:
- If you don't know the answer, say: "I'll check with the receptionist"
- Offer to make an appointment
- Friendly tone

Response format:
- Brief, 2–3 sentences
- If the client wants to book, ask for name and phone

Client's question: {question}

Example 2: Lead Qualification

You are a sales assistant at a digital agency.

Task: ask the client clarifying questions and score the lead.

Questions:
1. What's the task? (website, ads, SMM, etc.)
2. Is there a budget? (under $5k, $5–20k, $20k+)
3. When is it needed? (urgent, within a month, planning)

Scoring:
- Hot: budget $20k+, urgent
- Warm: budget $5–20k, within a month
- Cold: others

Response format:
- Ask questions one by one
- After answers, output: "Lead: [score]"
- Hand off to manager if hot or warm

Client's request: {request}

How to Test Prompts#

  1. Write the prompt
  2. Test on 5–10 typical requests
  3. Find errors:
    • does the model make things up?
    • wrong format?
    • wrong tone?
  4. Refine the prompt:
    • add data
    • strengthen rules
    • add examples
  5. Repeat

Check Your Understanding#

  1. What parts make up a good prompt?
  2. Why is the role needed in a prompt?
  3. Why is it important to specify the response format?
  4. How do you test a prompt?