Advanced Prompt Engineering Techniques

Why Basic Prompting Is Not Enough

For complex reasoning, code generation, and nuanced judgment, you need structured approaches that guide how the model thinks, not just what it produces.

Technique 1: Chain-of-Thought (CoT)

Force the model to show reasoning, which dramatically improves accuracy on multi-step problems.

Simple: "What is 15% of 847?"
Answer: 127.05
With CoT: "What is 15% of 847? Think step by step."
10% of 847 = 84.7
5% = 84.7 / 2 = 42.35
15% = 84.7 + 42.35 = 127.05
Answer: 127.05

CoT improves accuracy by 10-40% on math and logical reasoning tasks.

Technique 2: Self-Consistency Sampling

Generate 5+ CoT responses, take the majority answer. Improves accuracy significantly.

python
from collections import Counter
def self_consistent_query(question: str, samples: int = 5) -> str:
    prompt = f"Question: {question}\nThink step by step. Final answer on last line: Answer: X"
    answers = []
    for _ in range(samples):
        r = client.chat.completions.create(
            model='gpt-4o', messages=[{'role': 'user', 'content': prompt}],
            temperature=0.7
        )
        text = r.choices[0].message.content
        for line in text.split('\n'):
            if line.startswith('Answer:'):
                answers.append(line.replace('Answer:', '').strip())
    return Counter(answers).most_common(1)[0][0] if answers else 'Unknown'

Technique 3: Tree of Thoughts

Explore 3 approaches, score each, complete the best one:

Problem: {problem}
Generate 3 different approaches. For each:
Describe strategy briefly
Execute first few steps
Rate promise 1-10 with reason
Then select the best approach and complete it.

Technique 4: Constitutional AI Prompting

Define principles, then have the model critique and revise its own output:

You are a medical information assistant. Follow these principles:
ACCURACY: Only state scientifically established facts; flag uncertainty
SAFETY: Always recommend consulting a doctor for diagnosis/treatment
CLARITY: Use plain language for non-specialists
Draft an initial response, critique it against each principle, then revise.

Technique 5: XML-Tagged Structured Output

More reliable than asking for JSON directly:

Analyze this code and provide:

  
    
      critical|major|minor
      What is wrong
      Suggested fix
    
  
  1-10
  approve|request_changes

Technique 6: Specific Role Prompting

Weak: "You are a helpful assistant."

Strong: "You are a senior TypeScript engineer with 8 years building large-scale React applications. You follow functional programming principles, prioritize type safety, and write maintainable code. You are direct and critical - you point out problems rather than just agreeing."

Technique 7: Negative Space Prompting

Tell the model what NOT to do:

Write a marketing email.
Do NOT:
Use cliches like game-changing or revolutionary
Open with a question (Are you tired of...?)
Use passive voice
Make vague promises
Exceed 150 words
DO:
Lead with a specific problem
Include one concrete data point
End with a single clear call to action

Measuring Prompt Quality

Always evaluate systematically:

Create 20+ representative test cases

Define clear success criteria

Score outputs consistently

Track improvements version by version

Regression test before shipping new prompts