Prompting tips · AI Marketing

Prompting tips

**Prompt Engineering Insights from Sander Schulhoff and Team**

@SanderSchulhoff and his team analyzed over 1,500 academic papers spanning more than 200 prompting techniques.

Here are the **top 5 most effective prompt strategies**, along with practical examples:

### 1. **Few-Shot Prompting**

Instead of giving the model just a single query, include 2–5 examples to illustrate the desired output. This helps set expectations and improves accuracy.

**Example – Ticket Classification:**

```

Ticket: “I can’t reset my password.”

Category: Account Access

---

Ticket: “The app crashes when I upload a file.”

Category: Bug Report

---

Ticket: “I want to upgrade to the Pro plan.”

Category:

```

### 2. **Decomposition**

Break down complex problems into smaller steps before asking for a final answer.

**Example Prompt:**

> “What are the subproblems involved in this task?”

> ...

> “Great. Now solve each subproblem one at a time.”

### 3. **Self-Critique / Reflection**

Ask the model to evaluate or critique its own response. This can significantly enhance reasoning and reduce errors—especially in logical or high-stakes tasks.

**Example Prompts:**

> “Can you review your previous answer for accuracy?”

> “What would you change or improve about your response?”

> ...

> “Nice insight. Please apply that improvement now.”

### 4. **Add Rich Context**

The more relevant context you provide, the better the model performs—just like giving clear instructions to a human assistant.

**Less Effective:**

> *Craft a response to this ticket: “Why did my payment get declined? This is ridiculous.”*

**More Effective:**

> *You’re replying to a frustrated customer whose payment just failed. They’re on a paid plan. Keep the tone empathetic, avoid upselling, and include a link to retry payment.*

> *Here’s their message: “Why did my payment get declined? This is ridiculous.”*

### 5. **Ensemble Prompting**

Language models can vary in output. Try multiple prompt variations and compare results. Then:

* Rank the responses

* Vote on the best one

* Optionally, refine the top choice by refeeding it to the model

These techniques aren't just academic, they're practical, proven ways to get more reliable, intelligent, and useful responses from language models.

How about you? Any good tips that you use to get results from AI ?

0 comments