🎯 Learning Objectives
- Understand the difference between a question and a prompt
- Learn how LLMs predict rather than understand
- Identify why prompts fail and how to fix them
- Distinguish casual prompt writing from professional prompt engineering
- Bust the 5 most common myths about AI prompting
1. What Is a Prompt (Beyond 'a Question')
Most people treat prompts like Google searches — short, keyword-driven, hoping the AI figures it out. That works for trivial tasks and fails for professional work.
A prompt is just a question. Example: "Explain GMP." — like walking into a factory saying "Tell me about quality." You'll get something, but never what you actually need.
A prompt is instructions + context + expectations.
What Changed?
- Assigned a role — pharmaceutical quality consultant
- Defined the audience — supervisor with shop-floor experience
- Controlled depth — practical, not regulatory
- Shaped the output — 5 bullet points maximum
2. How LLMs Think: Tokens, Probability & Context Windows
LLMs are prediction engines, not knowledge databases. When you type "The capital of France is…" your phone predicts "Paris". LLMs work the same way — at massive scale.
AI does not know the answer. It predicts the most likely next token given everything you've told it. Clear, precise prompts produce better predictions.
Tokens
- "validation" — leads in one direction
- "process validation" — significantly different output path
- "cleaning validation in injectables" — very specific, very different result
Context Window: Order Matters
3. Why Prompting Works (and When It Fails)
Imagine telling a colleague: "Prepare a report." vs "Prepare a 1-page summary for the CEO highlighting key risks and recommendations." AI responds identically to this dynamic.
The AI doesn't know which instruction to prioritise. Result: generic, diluted output that satisfies nothing.
Common Failure Modes
- Vague task — "Write something about quality"
- No audience — depth and tone become guesses
- Conflicting constraints — diluted, compromised output
- No format specified — unstructured text that's hard to use immediately
4. Prompt Engineering vs Prompt Writing
| Aspect | Prompt Writing | Prompt Engineering |
|---|---|---|
| Approach | Conversational, ad-hoc | Deliberate, structured |
| Quality | Variable, inconsistent | Predictable, repeatable |
| Iteration | Random retries | Systematic refinement |
| Reusability | One-off | Templates & libraries |
| Analogy | Talking | Briefing |
Prompt writing is hoping. Prompt engineering is designing. One is a conversation — the other is a specification.
5. Common Myths and Misconceptions
Myth 1: Longer Prompts Are Always Better
3 vague paragraphs with no role, format, or audience.
Myth 2: AI Understands Intent
"Make this strong" — Strong how? Aggressive? Persuasive? Technical? The model will guess. Often wrongly. Never assume the AI knows what you mean.
Myth 3: Fancy Language Improves Output
Myth 4: One Perfect Prompt Exists
- First output → 70% right
- After one refinement → 85% right
- After critique loop → 95% usable
Prompting is iteration, not magic.
✏️ Module 01 Exercise
Take any topic from your daily pharma work. Write the worst possible one-sentence prompt. Then rebuild it by adding: (1) a role, (2) a defined audience, (3) depth constraints, and (4) an output format. Compare both outputs side by side.
🔑 Key Takeaways
- AI predicts — it does not understand
- The clearer your instructions, the better the prediction
- A prompt is a brief, not a question
- Conflicting instructions produce diluted output
- Prompting is iteration — expect to refine
- Better models reward better prompts exponentially more