Day 3 - AI Engineering Journey- Prompt Engineering Is Not What You Think

Day 3 — Prompt Engineering Is Not What You Think

AI Engineering — Day by Day
My journey to becoming an AI Engineer

---

When I first heard about prompt engineering, I thought it was about:

• Writing “smart” prompts
• Using fancy tricks
• Memorizing templates

But after actually digging into it, I realized something important:

Prompt engineering is not about clever wording.
It’s about reducing ambiguity for a probabilistic system.

This shift changed everything for me.

---

🧠 What I Understood Today

At its core, an LLM:

• Doesn’t “understand” like humans
• Doesn’t “know” the right answer

It just predicts the next most probable token

So, if your prompt is vague →
The model has too many possible directions →
And your output becomes inconsistent

---

❌ Why Most Prompts Fail

Explain trading

What’s wrong here?

• No context
• No audience
• No structure
• No constraints

The model is forced to guess what you want.

---

✅ Improving the Same Prompt

Explain stock market trading in 3 bullet points for a beginner using simple language

Now:

• Clear audience
• Clear format
• Clear expectation

👉 Output becomes more predictable and useful

---

🧩 The Structure I Learned

A good prompt is not random. It usually has:

[ROLE]

[TASK]

[CONSTRAINTS]

[OUTPUT FORMAT]

Example:

You are a financial analyst.

Explain options trading.

- Keep it under 100 words
- Use simple language
- Give one real-world example

Return answer in bullet points.

---

🔥 The Biggest Insight

The quality of output depends on how much you constrain the problem

More freedom for the model = more randomness
More constraints = more control

---

💭 Questions That Came to My Mind (And What I Learned)

While learning this, I had a few doubts. Writing them down actually helped me understand better.

---

❓ 1. Why do constraints improve output quality?

At first, I thought constraints just reduce verbosity.

But the real reason is deeper:

Constraints reduce the solution space the model has to explore.

Without constraints:

• Too many possible outputs
• More randomness

With constraints:

• Narrower possibilities
• More focused predictions

---

❓ 2. Is role prompting really necessary?

I wasn’t sure about this.

What I understood:

Role prompting is helpful, but not mandatory.

It works because:

• It biases the model toward a certain tone or domain

But:

• The model can still infer context from the task itself

So it’s more like a soft guide, not a requirement.

---

❓ 3. Why does breaking tasks into steps improve results?

Initially, I thought it just “adds clarity.”

But the actual reason is:

It reduces complexity by guiding the model through smaller steps.

Instead of solving everything at once:

• The model solves step-by-step
• Each step improves the next

This is why techniques like “think step by step” work.

---

🧠 Final Mental Model

After today, this is how I think about prompts:

A prompt is not a question — it’s a system design problem

You are:

• Defining input structure
• Reducing ambiguity
• Controlling output behavior

---

🚀 What Changed for Me

Before:

• I wrote prompts randomly
• Blamed the model when output was bad

Now:

• I see prompts as interfaces to a probabilistic system
• If output is bad → input design is probably bad

---

💭 Final Thought

LLMs are not unpredictable.

They just follow rules most people don’t understand.

Once you start designing prompts instead of guessing them —
you stop struggling… and start controlling the output.

---

This is Day 3 of my AI engineering journey — and this was a big shift in thinking.