Module 03: Prompting Patterns & Frameworks | Hello Pharma AI Upskilling Program

🎯 Learning Objectives

Distinguish zero-shot, one-shot, and few-shot prompting
Apply chain-of-thought correctly and know when not to
Use decomposition for complex tasks
Simulate expert perspectives with role-based prompting
Build critique–refine–rewrite loops for 2–3× quality improvement

1. Zero-Shot, One-Shot & Few-Shot Prompting

Zero-Shot

Use When

Task is simple. Output is generic. Style consistency doesn't matter.

"Summarise this paragraph."

One-Shot

Example: Input: Explain GMP to operators. Output: GMP ensures every shop-floor activity is controlled and documented. Now: Explain data integrity to operators.

Few-Shot

Example 1 — Explain validation → "Validation ensures processes consistently meet quality standards." Example 2 — Explain calibration → "Calibration ensures equipment gives accurate, reliable readings." Now: Explain cleaning validation.

Technique	Best For	Risk
Zero-Shot	Simple, generic tasks	High variance
One-Shot	Style anchoring	One bad example poisons output
Few-Shot	Branded, consistent series	More setup time

2. Chain-of-Thought — When to Use, When NOT To

✅ Use CoT For

Analysis, decision-making, complex reasoning, diagnosis, debugging.

Explain step by step why poor cleaning validation can lead to regulatory observations.

❌ Do NOT Use CoT For

Final customer-facing content, polished writing, short summaries, executive communications. CoT output sounds mechanical and verbose.

💡 Two-Step Approach

Step 1: Use CoT internally — "Think step by step about..."
Step 2: "Now write the final answer in polished language, without showing intermediate reasoning."

3. Decomposition — Break Complex Tasks

❌ Without Decomposition

"Create a complete pharma marketing strategy."

Overwhelming, shallow response that covers everything superficially.

✅ With Decomposition

Break this task into: 1. Market understanding 2. Target audience definition 3. Messaging strategy 4. Channel plan Solve each part first, then combine into a final strategy.

4. Role-Based Prompting (Expert Simulation)

❌ No Role

"Review this SOP."

✅ With Role

You are a regulatory auditor reviewing this SOP during an inspection. Identify potential gaps or red flags.

High-Value Roles for Pharma Work

Regulatory inspector — finds compliance gaps from an auditor's mindset
CXO advisor — executive framing and strategic implications
Devil's advocate — challenges assumptions, surfaces weaknesses
Patient/customer — evaluates from the end-user's perspective

5. Critique–Refine–Rewrite Loops

This is where quality jumps 2–3×. Run a deliberate editorial cycle instead of accepting the first output.

Step 1: Generate Draft an explanation of process validation. Step 2: Critique Critically review the above response. Identify unclear areas, logical gaps, or missing points. Do not rewrite yet — only diagnose. Step 3: Rewrite Rewrite addressing every gap while keeping it concise and professional.

💡 Why This Works

The model is better at critiquing its own output than at producing perfection in one attempt. Separating generation from evaluation unlocks a quality level that single-pass prompting can never reach.

✏️ Module 03 Exercise

Pick a complex task from your pharma role. Write a single-shot prompt and note quality. Then decompose it into 3–5 sub-tasks and run each. Finally apply the critique loop to the weakest section. Track the quality at each stage.

🔑 Key Takeaways

Few-shot examples teach style better than explicit style instructions
Chain-of-thought is for reasoning, not for polished final outputs
Decompose complexity before you attempt to solve it
Roles shape perspective and reasoning — not just tone
Structured comparison criteria produce decision-ready outputs
The critique–refine loop consistently produces 2–3× quality improvement

← M02: Anatomy of a High-Performanc… M04: Controlling Output Quality →