Prompt Engineering: Best Practices and Anti-Patterns

The difference between a mediocre AI response and an excellent one is often just the prompt. Good prompting is a skill that directly impacts the quality of your AI system.

Core Principles

1. Clarity Over Cuteness

Bad:

"yo, can u help me with some code? 🤔"

Good:

"Write a Python function that validates email addresses using regex"

2. Specificity Beats Vagueness

Bad:

"Explain AI"

Good:

"Explain how transformer models use attention mechanisms to process text sequences"

3. Context Matters

Bad:

"Is this good?"

Good:

"Review this customer service response for tone, accuracy, and helpfulness:
[response text]"

Structural Best Practices

System vs. User Messages

Keep instructions and data separate:

System Message (your rules):
"You are a Python expert. Answer only Python questions."

User Message (the actual request):
"Write a function to reverse a string"

Format Your Input

Use clear sections:

CONTEXT:
[background information]

TASK:
[what you want done]

CONSTRAINTS:
[limitations or requirements]

FORMAT:
[how you want the output]

Role-Based Prompting

Assign the AI a role:

GOOD: "You are an experienced product manager. Analyze this feature request and identify potential issues"

VS.

BAD: "Analyze this feature request"

Why: The first prompt activates relevant expertise patterns

Output Control Techniques

1. Specify Output Format

"Respond in JSON format:
{
  'action': 'string',
  'priority': 'high/medium/low',
  'reasoning': 'string'
}"

2. Length Constraints

"In 100 words or less: ..."
"Keep your answer to 2-3 sentences: ..."
"Write an exhaustive 5-paragraph essay: ..."

Specific lengths > generic “be concise”

3. Example Output

"Output should look like:
Status: [good/bad]
Risk Level: [1-10]
Next Steps: [list]"

Prompt Patterns

The Persona Pattern

"You are a [expert role] with [specific expertise].
Your communication style is [style].
You [specific trait or capability].

Now, [task]"

Example:

"You are a security expert specializing in cloud infrastructure.
Your communication style is direct and technical.
You identify risks clearly and provide actionable mitigations.

Now, review this AWS configuration for security issues."

The Few-Shot Pattern

"Here are examples of [task]:
[Example 1]
[Example 2]
[Example 3]

Now, [new task similar to examples]"

The Chain-of-Thought Pattern

"Solve this step by step:
1. [first step]
2. [second step]
3. [final step]

Problem: [task]"

Or even simpler:

"Solve this problem step by step:
[problem]"

(Model will naturally break it down)

The Constraint Pattern

"Given these constraints:
- [constraint 1]
- [constraint 2]
- [constraint 3]

Complete this task: [task]"

Temperature and Randomness

For prompts:

• Deterministic task (extraction, classification): Temperature 0-0.3
• Balanced task (general Q&A): Temperature 0.7-0.9
• Creative task (brainstorming): Temperature 1.0-1.2

Don’t use high temperature for fact-based tasks.

Anti-Patterns (Things to Avoid)

❌ Being Too Polite

BAD: "If you don't mind, could you possibly help with...?"
GOOD: "Summarize this document"

Why: AI isn't offended; extra politeness wastes tokens

❌ Apologizing Unnecessarily

BAD: "Sorry to bother you, but I have a complex question..."
GOOD: "[Ask the complex question directly]"

Why: Again, wastes tokens and adds nothing

❌ Weak Constraints

BAD: "Try to be concise"
GOOD: "In 150 words or less"

Why: Specific constraints work; vague suggestions don't

❌ Contradictory Instructions

BAD: "Be creative but accurate", "Think outside the box but follow all guidelines"
GOOD: "Prioritize accuracy. You may suggest creative alternatives if they're factually sound."

Why: Resolve conflicts explicitly

❌ Expecting Common Sense

BAD: "Assume you know the context"
GOOD: "[Provide explicit context]"

Why: AI can't read minds; provide what you think is obvious

❌ Vague Success Criteria

BAD: "Write good code"
GOOD: "Write efficient Python code that handles edge cases and follows PEP 8"

Why: Concrete criteria enable better output

Advanced Techniques

Negative Examples

"Here's what NOT to do:
[example of bad output]

Now do this correctly:
[task]"

Why: Negative examples help model understand boundaries.

Decomposition

Instead of:
"Build a complete e-commerce system"

Use:
"First, design the database schema for an e-commerce system.
Then, write the API endpoints for product management."

Why: Breaks complex tasks into manageable pieces

Authority Grounding

"According to [authoritative source], ..."
"Based on industry best practices, ..."
"Following the official documentation, ..."

Why: Anchors model in credible information

Hypothesis Testing

"I believe [hypothesis]. Is this correct?
Here's my reasoning: [reasoning]"

Why: Model corrects incorrect assumptions more effectively

Testing Your Prompts

1. A/B Testing

Version A: Generic prompt
Version B: Structured prompt
Version C: Few-shot prompt

Test on 10-20 examples, measure quality
Pick the best version

2. Adversarial Testing

Try to break your prompt:
- What if input is ambiguous?
- What if input is very long?
- What if input is very short?
- What if input is in a different language?

Refine prompt based on failures

3. Consistency Testing

Ask the same question 5 times
Do you get similar answers?
If not, your prompt is under-specified

Prompt Optimization Checklist

• Is my task clearly defined?
• Did I provide necessary context?
• Did I specify the output format?
• Are my constraints specific and clear?
• Did I remove unnecessary politeness?
• Did I resolve any conflicting instructions?
• Is my prompt using the right temperature?
• Have I tested on representative examples?
• Does the output meet my success criteria?
• Are there edge cases I haven’t considered?

Real-World Examples

Example 1: Code Review

Before:

"Review this code"

After:

"Review this Python code for:
1. Security vulnerabilities
2. Performance issues
3. Adherence to PEP 8
4. Potential bugs or edge cases

Output format: 
- Issue: [description]
- Severity: [high/medium/low]
- Fix: [suggested fix]

Code:
[code]"

Example 2: Content Classification

Before:

"Is this positive or negative?"

After:

"Classify this customer review as:
- Positive (customer satisfied)
- Negative (customer unsatisfied)
- Neutral (no clear sentiment)
- Mixed (both positive and negative elements)

Also provide:
- Confidence: [0-100]%
- Key sentiment words: [list]
- Suggested response if needed: [optional]

Review: [review text]"

The Golden Rule

Treat the AI like a smart but new employee:

• Give explicit instructions (don’t assume they know)
• Provide examples (show what good looks like)
• Define success criteria (how will you judge success?)
• Provide context (what problem are we solving?)