10 Practice Lab: Master Prompt Engineering
10.1 What You’ll Build
A Prompt Benchmarking Toolkit that tests multiple prompt strategies against the same task and scores them.
10.2 Step 1: Prompt Comparison Engine
# file: prompt_benchmark.py
from openai import OpenAI
from dataclasses import dataclass
from typing import Callable
import json
client = OpenAI()
@dataclass
class PromptResult:
technique: str
prompt: str
response: str
tokens_used: int
def test_prompt(technique: str, messages: list) -> PromptResult:
"""Run a single prompt and capture results."""
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=messages,
temperature=0.3
)
return PromptResult(
technique=technique,
prompt=str(messages),
response=response.choices[0].message.content,
tokens_used=response.usage.total_tokens
)
# Define the task
TASK = "A customer says: 'Your app crashed and I lost all my data. Fix it NOW!'"
# Test all techniques
prompts_to_test = [
("Zero-Shot", [
{"role": "user", "content": f"Respond to this customer complaint: {TASK}"}
]),
("Role + Context", [
{"role": "system", "content": "You are a senior customer success manager at a tech startup. You are empathetic, solution-focused, and professional."},
{"role": "user", "content": f"Draft a response to: {TASK}"}
]),
("Chain-of-Thought", [
{"role": "user", "content": f"""
Draft a customer service response to this complaint.
Think through this step by step:
1. What is the customer's emotional state?
2. What is the core problem?
3. What immediate action can we take?
4. What is the ideal tone?
Then write the response.
Complaint: {TASK}
"""}
]),
("Few-Shot", [
{"role": "user", "content": f"""
Here are examples of excellent customer service responses:
Complaint: "The delivery was late!"
Response: "I sincerely apologise for the delayed delivery. I've flagged this as urgent and your order will be prioritised. You'll receive a 10% discount on your next order as our apology."
Complaint: "I was charged twice for my subscription."
Response: "I'm so sorry for this billing error — that should never happen. I've immediately reversed the duplicate charge and you'll see the refund within 3-5 business days."
Now respond to this complaint using the same empathetic, action-oriented style:
{TASK}
"""}
])
]
results = []
for technique, messages in prompts_to_test:
result = test_prompt(technique, messages)
results.append(result)
print(f"\n{'='*60}")
print(f"Technique: {result.technique} ({result.tokens_used} tokens)")
print(f"Response:\n{result.response}")
# Save results for analysis
with open("benchmark_results.json", "w") as f:
json.dump([vars(r) for r in results], f, indent=2)10.3 Step 2: Automated Prompt Evaluator
Use an LLM to score prompts against criteria:
def evaluate_response(response: str, criteria: list[str]) -> dict:
"""Use GPT-4o to evaluate a response against criteria."""
eval_prompt = f"""
Evaluate this customer service response against each criterion.
Score each from 1-10 and give a one-line reason.
Response to evaluate:
"{response}"
Criteria:
{json.dumps(criteria, indent=2)}
Return JSON only:
{{
"criteria_name": {{"score": X, "reason": "..."}},
...
"overall_score": X
}}
"""
eval_response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": eval_prompt}],
response_format={"type": "json_object"},
temperature=0
)
return json.loads(eval_response.choices[0].message.content)
# Evaluate all results
criteria = ["Empathy", "Actionability", "Professionalism", "Clarity"]
print("\n📊 EVALUATION SCORES")
print("=" * 60)
for result in results:
scores = evaluate_response(result.response, criteria)
print(f"\n{result.technique}:")
for criterion in criteria:
s = scores.get(criterion, {})
print(f" {criterion}: {s.get('score', 'N/A')}/10 — {s.get('reason', '')}")
print(f" Overall: {scores.get('overall_score', 'N/A')}/10")10.4 Step 3: Build a Prompt Template Library
# file: prompt_library.py
from string import Template
from dataclasses import dataclass
@dataclass
class PromptTemplate:
name: str
description: str
template: str
required_vars: list[str]
def fill(self, **kwargs) -> str:
t = Template(self.template)
return t.safe_substitute(**kwargs)
PROMPT_LIBRARY = {
"executive_summary": PromptTemplate(
name="Executive Summary",
description="Create a concise executive summary",
template="""You are a senior business analyst.
Create a 3-bullet executive summary of the following content.
Each bullet should be one clear sentence.
Audience: ${audience}
Content: ${content}
Executive Summary:""",
required_vars=["audience", "content"]
),
"swot_analysis": PromptTemplate(
name="SWOT Analysis",
description="Generate a SWOT analysis",
template="""You are a strategic business consultant.
Analyse ${company} in the ${industry} industry.
Provide a concise SWOT analysis with 3 points per category.
Format as:
STRENGTHS:
- ...
WEAKNESSES:
- ...
OPPORTUNITIES:
- ...
THREATS:
- ...""",
required_vars=["company", "industry"]
),
"code_review": PromptTemplate(
name="Code Review",
description="Review code for issues and improvements",
template="""You are a senior ${language} developer.
Review the following code for:
1. Bugs and errors
2. Security issues
3. Performance improvements
4. Code style and best practices
Code:
```${language}
${code}Provide specific, actionable feedback:“““, required_vars=[“language”, “code”] ) }
11 Usage example
summary_prompt = PROMPT_LIBRARY[“executive_summary”].fill( audience=“C-suite executives”, content=“Detailed report text here…” ) print(summary_prompt)
---
## Lab Challenges 🏆
1. **Easy**: Add 3 more prompt templates to the library for your specific use case
2. **Medium**: Build a Streamlit or Gradio UI for the prompt comparison tool
3. **Hard**: Implement automatic prompt optimisation — use the evaluator scores to iteratively improve a prompt
---
::: {.callout-note icon="false"}
## ✅ Lab Complete!
You've mastered the fundamentals of prompt engineering. Next up: **Vector Databases** in Chapter 9.
:::
:::{#quarto-navigation-envelope .hidden}
[Fundamentals of AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyLXRpdGxl"}
[Fundamentals of AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1uYXZiYXItdGl0bGU="}
[<span class='chapter-number'>11</span> <span class='chapter-title'>Vector Databases Deep Dive</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1uZXh0"}
[<span class='chapter-number'>9</span> <span class='chapter-title'>Prompt Engineering Techniques</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1wcmV2"}
[<span class='chapter-number'>1</span> <span class='chapter-title'>cover.html</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jb3Zlci5odG1sPHNwYW4tY2xhc3M9J2NoYXB0ZXItbnVtYmVyJz4xPC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPmNvdmVyLmh0bWw8L3NwYW4+"}
[Preface]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9pbmRleC5odG1sUHJlZmFjZQ=="}
[Foundations — The Genesis of Modern AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tMQ=="}
[<span class='chapter-number'>2</span> <span class='chapter-title'>From Synapses to Silicon: The Genesis of Modern AI</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wMC1nZW5lc2lzLW9mLWFpLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjI8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+RnJvbS1TeW5hcHNlcy10by1TaWxpY29uOi1UaGUtR2VuZXNpcy1vZi1Nb2Rlcm4tQUk8L3NwYW4+"}
[Part I — Understanding AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tMg=="}
[<span class='chapter-number'>3</span> <span class='chapter-title'>Introduction to AI Agents</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wMS1pbnRyby1haS1hZ2VudHMuaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+Mzwvc3Bhbj4tLTxzcGFuLWNsYXNzPSdjaGFwdGVyLXRpdGxlJz5JbnRyb2R1Y3Rpb24tdG8tQUktQWdlbnRzPC9zcGFuPg=="}
[<span class='chapter-number'>4</span> <span class='chapter-title'>How LLMs Work in Real Time</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wMi1ob3ctbGxtcy13b3JrLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjQ8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+SG93LUxMTXMtV29yay1pbi1SZWFsLVRpbWU8L3NwYW4+"}
[<span class='chapter-number'>5</span> <span class='chapter-title'>Embeddings & Vector Representations</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wMy1lbWJlZGRpbmdzLXZlY3RvcnMuaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+NTwvc3Bhbj4tLTxzcGFuLWNsYXNzPSdjaGFwdGVyLXRpdGxlJz5FbWJlZGRpbmdzLSYtVmVjdG9yLVJlcHJlc2VudGF0aW9uczwvc3Bhbj4="}
[<span class='chapter-number'>6</span> <span class='chapter-title'>How LangChain Works</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wNC1ob3ctbGFuZ2NoYWluLXdvcmtzLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjY8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+SG93LUxhbmdDaGFpbi1Xb3Jrczwvc3Bhbj4="}
[Part II — Your First Hands-On Labs]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tMw=="}
[<span class='chapter-number'>7</span> <span class='chapter-title'>Practice Lab: Your First AI API Call</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wNS1sYWItZmlyc3QtYXBpLWNhbGwuaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+Nzwvc3Bhbj4tLTxzcGFuLWNsYXNzPSdjaGFwdGVyLXRpdGxlJz5QcmFjdGljZS1MYWI6LVlvdXItRmlyc3QtQUktQVBJLUNhbGw8L3NwYW4+"}
[<span class='chapter-number'>8</span> <span class='chapter-title'>Practice Lab: LangChain</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wNi1sYWItbGFuZ2NoYWluLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjg8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+UHJhY3RpY2UtTGFiOi1MYW5nQ2hhaW48L3NwYW4+"}
[Part III — Prompt Engineering]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tNA=="}
[<span class='chapter-number'>9</span> <span class='chapter-title'>Prompt Engineering Techniques</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wNy1wcm9tcHQtZW5naW5lZXJpbmcuaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+OTwvc3Bhbj4tLTxzcGFuLWNsYXNzPSdjaGFwdGVyLXRpdGxlJz5Qcm9tcHQtRW5naW5lZXJpbmctVGVjaG5pcXVlczwvc3Bhbj4="}
[<span class='chapter-number'>10</span> <span class='chapter-title'>Practice Lab: Master Prompt Engineering</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wOC1sYWItcHJvbXB0LWVuZ2luZWVyaW5nLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjEwPC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPlByYWN0aWNlLUxhYjotTWFzdGVyLVByb21wdC1FbmdpbmVlcmluZzwvc3Bhbj4="}
[Part IV — Vector Databases & Semantic Search]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tNQ=="}
[<span class='chapter-number'>11</span> <span class='chapter-title'>Vector Databases Deep Dive</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8wOS12ZWN0b3ItZGF0YWJhc2VzLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjExPC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPlZlY3Rvci1EYXRhYmFzZXMtRGVlcC1EaXZlPC9zcGFuPg=="}
[<span class='chapter-number'>12</span> <span class='chapter-title'>Practice Lab: Build a Semantic Search Engine</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xMC1sYWItc2VtYW50aWMtc2VhcmNoLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjEyPC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPlByYWN0aWNlLUxhYjotQnVpbGQtYS1TZW1hbnRpYy1TZWFyY2gtRW5naW5lPC9zcGFuPg=="}
[Part V — Retrieval Augmented Generation]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tNg=="}
[<span class='chapter-number'>13</span> <span class='chapter-title'>RAG: Retrieval Augmented Generation</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xMS1yYWcuaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+MTM8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+UkFHOi1SZXRyaWV2YWwtQXVnbWVudGVkLUdlbmVyYXRpb248L3NwYW4+"}
[<span class='chapter-number'>14</span> <span class='chapter-title'>Practice Lab: RAG Implementation</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xMi1sYWItcmFnLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjE0PC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPlByYWN0aWNlLUxhYjotUkFHLUltcGxlbWVudGF0aW9uPC9zcGFuPg=="}
[Part VI — Advanced AI Workflows]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOnF1YXJ0by1zaWRlYmFyLXNlY3Rpb24tNw=="}
[<span class='chapter-number'>15</span> <span class='chapter-title'>LangGraph for AI Workflows</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xMy1sYW5nZ3JhcGguaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+MTU8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+TGFuZ0dyYXBoLWZvci1BSS1Xb3JrZmxvd3M8L3NwYW4+"}
[<span class='chapter-number'>16</span> <span class='chapter-title'>Practice Lab: Build a Stateful AI Workflow</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xNC1sYWItbGFuZ2dyYXBoLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjE2PC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPlByYWN0aWNlLUxhYjotQnVpbGQtYS1TdGF0ZWZ1bC1BSS1Xb3JrZmxvdzwvc3Bhbj4="}
[<span class='chapter-number'>17</span> <span class='chapter-title'>Model Context Protocol (MCP)</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xNS1tY3AuaHRtbDxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+MTc8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+TW9kZWwtQ29udGV4dC1Qcm90b2NvbC0oTUNQKTwvc3Bhbj4="}
[<span class='chapter-number'>18</span> <span class='chapter-title'>Practice Lab: Advanced MCP Concepts</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xNi1sYWItbWNwLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjE4PC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPlByYWN0aWNlLUxhYjotQWR2YW5jZWQtTUNQLUNvbmNlcHRzPC9zcGFuPg=="}
[<span class='chapter-number'>19</span> <span class='chapter-title'>Conclusion: Your AI Journey Continues</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9jaGFwdGVycy8xNy1jb25jbHVzaW9uLmh0bWw8c3Bhbi1jbGFzcz0nY2hhcHRlci1udW1iZXInPjE5PC9zcGFuPi0tPHNwYW4tY2xhc3M9J2NoYXB0ZXItdGl0bGUnPkNvbmNsdXNpb246LVlvdXItQUktSm91cm5leS1Db250aW51ZXM8L3NwYW4+"}
[References]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWludC1zaWRlYmFyOi9yZWZlcmVuY2VzLmh0bWxSZWZlcmVuY2Vz"}
[Part III — Prompt Engineering]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWJyZWFkY3J1bWJzLVBhcnQtSUlJLeKAlC1Qcm9tcHQtRW5naW5lZXJpbmc="}
[<span class='chapter-number'>10</span> <span class='chapter-title'>Practice Lab: Master Prompt Engineering</span>]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLWJyZWFkY3J1bWJzLTxzcGFuLWNsYXNzPSdjaGFwdGVyLW51bWJlcic+MTA8L3NwYW4+LS08c3Bhbi1jbGFzcz0nY2hhcHRlci10aXRsZSc+UHJhY3RpY2UtTGFiOi1NYXN0ZXItUHJvbXB0LUVuZ2luZWVyaW5nPC9zcGFuPg=="}
:::{.hidden .quarto-markdown-envelope-contents render-id="Zm9vdGVyLWxlZnQ="}
© 2026 Prof. Bongo Adi | Lagos Business School
:::
:::{.hidden .quarto-markdown-envelope-contents render-id="Zm9vdGVyLXJpZ2h0"}
Built with [Quarto](https://quarto.org)
:::
:::
:::{#quarto-meta-markdown .hidden}
[[[10]{.chapter-number} [Practice Lab: Master Prompt Engineering]{.chapter-title}]{#sec-lab-prompts .quarto-section-identifier} – Fundamentals of AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLW1ldGF0aXRsZQ=="}
[[[10]{.chapter-number} [Practice Lab: Master Prompt Engineering]{.chapter-title}]{#sec-lab-prompts .quarto-section-identifier} – Fundamentals of AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLXR3aXR0ZXJjYXJkdGl0bGU="}
[[[10]{.chapter-number} [Practice Lab: Master Prompt Engineering]{.chapter-title}]{#sec-lab-prompts .quarto-section-identifier} – Fundamentals of AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLW9nY2FyZHRpdGxl"}
[Fundamentals of AI]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLW1ldGFzaXRlbmFtZQ=="}
[A hands-on, interactive introduction to Artificial Intelligence for professionals
and students. From understanding how large language models work to building
production-ready RAG pipelines and AI workflows — no PhD required.
]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLXR3aXR0ZXJjYXJkZGVzYw=="}
[A hands-on, interactive introduction to Artificial Intelligence for professionals
and students. From understanding how large language models work to building
production-ready RAG pipelines and AI workflows — no PhD required.
]{.hidden .quarto-markdown-envelope-contents render-id="cXVhcnRvLW9nY2FyZGRkZXNj"}
:::
<!-- -->
::: {.quarto-embedded-source-code}
```````````````````{.markdown shortcodes="false"}
# Practice Lab: Master Prompt Engineering {#sec-lab-prompts}
::: {.callout-important icon="false"}
## 🧪 Lab Overview
**Duration:** 90–120 minutes | **Difficulty:** ⭐⭐☆☆☆
**Goal:** Build a prompt comparison tool and benchmark different techniques on real tasks.
:::
## What You'll Build
A **Prompt Benchmarking Toolkit** that tests multiple prompt strategies against the same task and scores them.
---
## Step 1: Prompt Comparison Engine
```python
# file: prompt_benchmark.py
from openai import OpenAI
from dataclasses import dataclass
from typing import Callable
import json
client = OpenAI()
@dataclass
class PromptResult:
technique: str
prompt: str
response: str
tokens_used: int
def test_prompt(technique: str, messages: list) -> PromptResult:
"""Run a single prompt and capture results."""
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=messages,
temperature=0.3
)
return PromptResult(
technique=technique,
prompt=str(messages),
response=response.choices[0].message.content,
tokens_used=response.usage.total_tokens
)
# Define the task
TASK = "A customer says: 'Your app crashed and I lost all my data. Fix it NOW!'"
# Test all techniques
prompts_to_test = [
("Zero-Shot", [
{"role": "user", "content": f"Respond to this customer complaint: {TASK}"}
]),
("Role + Context", [
{"role": "system", "content": "You are a senior customer success manager at a tech startup. You are empathetic, solution-focused, and professional."},
{"role": "user", "content": f"Draft a response to: {TASK}"}
]),
("Chain-of-Thought", [
{"role": "user", "content": f"""
Draft a customer service response to this complaint.
Think through this step by step:
1. What is the customer's emotional state?
2. What is the core problem?
3. What immediate action can we take?
4. What is the ideal tone?
Then write the response.
Complaint: {TASK}
"""}
]),
("Few-Shot", [
{"role": "user", "content": f"""
Here are examples of excellent customer service responses:
Complaint: "The delivery was late!"
Response: "I sincerely apologise for the delayed delivery. I've flagged this as urgent and your order will be prioritised. You'll receive a 10% discount on your next order as our apology."
Complaint: "I was charged twice for my subscription."
Response: "I'm so sorry for this billing error — that should never happen. I've immediately reversed the duplicate charge and you'll see the refund within 3-5 business days."
Now respond to this complaint using the same empathetic, action-oriented style:
{TASK}
"""}
])
]
results = []
for technique, messages in prompts_to_test:
result = test_prompt(technique, messages)
results.append(result)
print(f"\n{'='*60}")
print(f"Technique: {result.technique} ({result.tokens_used} tokens)")
print(f"Response:\n{result.response}")
# Save results for analysis
with open("benchmark_results.json", "w") as f:
json.dump([vars(r) for r in results], f, indent=2)11.1 Step 2: Automated Prompt Evaluator
Use an LLM to score prompts against criteria:
def evaluate_response(response: str, criteria: list[str]) -> dict:
"""Use GPT-4o to evaluate a response against criteria."""
eval_prompt = f"""
Evaluate this customer service response against each criterion.
Score each from 1-10 and give a one-line reason.
Response to evaluate:
"{response}"
Criteria:
{json.dumps(criteria, indent=2)}
Return JSON only:
{{
"criteria_name": {{"score": X, "reason": "..."}},
...
"overall_score": X
}}
"""
eval_response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": eval_prompt}],
response_format={"type": "json_object"},
temperature=0
)
return json.loads(eval_response.choices[0].message.content)
# Evaluate all results
criteria = ["Empathy", "Actionability", "Professionalism", "Clarity"]
print("\n📊 EVALUATION SCORES")
print("=" * 60)
for result in results:
scores = evaluate_response(result.response, criteria)
print(f"\n{result.technique}:")
for criterion in criteria:
s = scores.get(criterion, {})
print(f" {criterion}: {s.get('score', 'N/A')}/10 — {s.get('reason', '')}")
print(f" Overall: {scores.get('overall_score', 'N/A')}/10")11.2 Step 3: Build a Prompt Template Library
# file: prompt_library.py
from string import Template
from dataclasses import dataclass
@dataclass
class PromptTemplate:
name: str
description: str
template: str
required_vars: list[str]
def fill(self, **kwargs) -> str:
t = Template(self.template)
return t.safe_substitute(**kwargs)
PROMPT_LIBRARY = {
"executive_summary": PromptTemplate(
name="Executive Summary",
description="Create a concise executive summary",
template="""You are a senior business analyst.
Create a 3-bullet executive summary of the following content.
Each bullet should be one clear sentence.
Audience: ${audience}
Content: ${content}
Executive Summary:""",
required_vars=["audience", "content"]
),
"swot_analysis": PromptTemplate(
name="SWOT Analysis",
description="Generate a SWOT analysis",
template="""You are a strategic business consultant.
Analyse ${company} in the ${industry} industry.
Provide a concise SWOT analysis with 3 points per category.
Format as:
STRENGTHS:
- ...
WEAKNESSES:
- ...
OPPORTUNITIES:
- ...
THREATS:
- ...""",
required_vars=["company", "industry"]
),
"code_review": PromptTemplate(
name="Code Review",
description="Review code for issues and improvements",
template="""You are a senior ${language} developer.
Review the following code for:
1. Bugs and errors
2. Security issues
3. Performance improvements
4. Code style and best practices
Code:
```${language}
${code}Provide specific, actionable feedback:“““, required_vars=[“language”, “code”] ) }
12 Usage example
summary_prompt = PROMPT_LIBRARY[“executive_summary”].fill( audience=“C-suite executives”, content=“Detailed report text here…” ) print(summary_prompt)
---
## Lab Challenges 🏆
1. **Easy**: Add 3 more prompt templates to the library for your specific use case
2. **Medium**: Build a Streamlit or Gradio UI for the prompt comparison tool
3. **Hard**: Implement automatic prompt optimisation — use the evaluator scores to iteratively improve a prompt
---
::: {.callout-note icon="false"}
## ✅ Lab Complete!
You've mastered the fundamentals of prompt engineering. Next up: **Vector Databases** in Chapter 9.
::::::