Garbage In, Garbage Out

We have all heard the phrase. In Generative AI, it is the absolute law. If you give a model a vague, lazy instruction, you will get a vague, lazy result.

Prompt Engineering is often misunderstood as "hacks" or "magic words." It isn't. It is communication hygiene. It is the skill of clearly defining a task so that a machine (which lacks common sense) can execute it.

In this lesson, we will formalize the structure of a prompt. As a leader, you may not write every prompt, but you will review them. You need to be able to look at a failed AI output and say, "The prompt was missing Context."

---

1. The Anatomy of a Perfect Prompt

Effective prompts typically contain four key components. We can remember them with the acronym CO-STAR (Context, Output, Style, Task, Audience, Requirements) or simply the 4-Block Framework:

1. Context (The Setup)

Who is the AI? What is the background?

• Bad: "Write an email."
• Good: "You are an empathetic Senior Customer Support Agent for a luxury watch brand. A customer is angry because their delivery is late."

2. Task (The Instruction)

What, specifically, should the AI do? Use action verbs.

• Good: "Draft a response apologizing for the delay, explaining it was due to a snowstorm, and offering a 10% discount on their next repair."

3. Constraints (The Guardrails)

What should the AI NOT do? This is critical for business safety.

• Good: "Do not promise a specific delivery date. The email must be under 150 words. Do not use slang."

4. Output Format (The Container)

How do you want the data back?

• Good: "Format the output as a JSON object with fields: 'subject_line' and 'body_text'."

---

2. Iterative Prompting Strategy

No one writes the perfect prompt on the first try. Generative AI development is closer to gardening than construction. You plant a prompt, see what grows, prune it, and try again.

The Iteration Cycle

graph TD
    A[Draft Prompt v1] --> B{Run Test Case}
    B --> C[Analyze Output]
    C -->|Too Vague?| D[Add Context]
    C -->|Hallucinating?| E[Add Grounding/Constraints]
    C -->|Wrong Tone?| F[Adjust Persona]
    D --> G[Draft Prompt v2]
    E --> G
    F --> G
    G --> B
    C -->|Success| H[Deploy]
    
    style H fill:#34A853,stroke:#fff,stroke-width:2px,color:#fff

Real-World Example: Refining a Prompt

Goal: Summarize a meeting transcript.

Iteration 1 (Fail):

"Summarize this." Result: The model gives a long, rambling paragraph that misses the key decisions.

Iteration 2 (Better):

"You are a Project Manager. Summarize the meeting transcript below. Focus on action items." Result: Better, but it includes too much "chit chat" (e.g., "Bob asked about lunch").

Iteration 3 (Production Ready):

"You are a strict Project Manager. Analyze the transcript below.

1. List 'Decisions Made' in bullet points.
2. List 'Action Items' with the owner's name.
3. Ignore all small talk and pleasantries.
4. If no deadline is mentioned, mark it as 'TBD'. [Transcript Attached]"

---

3. Advanced Strategy: Few-Shot Prompting

We touched on this in Module 1, but let's see its business impact. Few-Shot Prompting is the single most effective way to align the model to your company's unique style without training a custom model.

Scenario: You want the AI to generate SQL queries from English questions.

Zero-Shot (Risky):

"Turn this into SQL: Show me top users." Result: SELECT * FROM users WHERE rank = 'top'; (This fails because your table is named customer_profiles, not users).

Few-Shot (Systematic):

"Turn English into SQL for our database. Table: customer_profiles (id, name, spend).

Example 1: Input: Show me top users. Output: SELECT name FROM customer_profiles ORDER BY spend DESC LIMIT 10;

Example 2: Input: Who is the newest customer? Output: SELECT name FROM customer_profiles ORDER BY created_at DESC LIMIT 1;

Input: Show me users who spent more than $100." Result: SELECT name FROM customer_profiles WHERE spend > 100;

By giving just two examples, you taught the model your schema and your intent.

---

4. Code Example: System Instructions

In Vertex AI (Gemini 1.5), the best practice is to put the "Context" and "Constraints" into the system_instruction field, and the specific task in the user prompt.

import vertexai
from vertexai.generative_models import GenerativeModel

# The "Permanent" behavior of the bot
system_prompt = """
You are a helpful IT Support assistant.
You only answer questions about password resets and wifi access.
If a user asks about anything else (like "Who won the game?"), politely decline.
Keep answers under 3 sentences.
"""

model = GenerativeModel(
    "gemini-1.5-flash",
    system_instruction=[system_prompt]
)

# The "Specific" interaction
user_question = "My internet isn't working."

response = model.generate_content(user_question)
print(response.text)

---

5. Summary

Prompt Engineering is the primary interface for programming modern AI.

1. Structure matters: Always include Context, Task, Constraints, and Output Format.
2. Iterate: Expect to rewrite your prompt 5-10 times using a test dataset.
3. Examples are power: Use Few-Shot prompting to define style and logic patterns.
4. System Instructions: Separate the "Who you are" (System) from the "What I want now" (User) for stability.

In the next lesson, we face the biggest problem in GenAI: The Knowledge Cutoff. We will learn about RAG (Retrieval Augmented Generation), the architecture that lets AI read your private data.

---

Knowledge Check