Prompt Engineering Best Practices: Get Better Results from Any LLM

Prompt engineering is the practice of structuring inputs to language models to get reliable, high-quality outputs. As LLMs have become more capable, effective prompting has become less about "tricks" and more about clear communication — but the techniques that consistently work are well established and worth knowing.

The Foundation: What LLMs Actually Do

Understanding prompt engineering starts with understanding what LLMs do: they predict the most probable continuation of the text they've been given. Your prompt is the context that shapes what "most probable" looks like.

A vague prompt produces a probable generic response. A specific prompt makes certain types of responses much more probable than others — and that's what good prompting achieves.

---

Technique 1: Specify Role, Goal, Format, and Constraints

The four-component prompt framework produces consistently better outputs:

Role: Who is the model playing? Establishing expertise helps. Goal: What outcome do you actually need? Format: What should the output look like? Constraints: What should be avoided or limited?

Weak prompt:

"Explain machine learning."

Strong prompt:

"You are a technical writer explaining machine learning to a business executive with no technical background. Write a 200-word explanation of what machine learning is and why companies use it. Avoid jargon; use one concrete business example. Do not use the words 'algorithm' or 'neural network.'"

The second prompt produces a usable, specific output. The first produces a textbook definition that may or may not serve your purpose.

---

Technique 2: Chain-of-Thought (CoT) Prompting

Chain-of-thought prompting instructs the model to reason step by step before producing its final answer. It dramatically improves accuracy on tasks requiring multiple logical steps.

How to use it:

• Add "Think step by step" before the question
• Or: "Reason through this carefully before giving your answer"
• Or: "Work through this problem step by step, showing your reasoning"

When to use CoT:

• Math problems and calculations
• Multi-step logical reasoning
• Code debugging ("explain what this code does step by step, then identify the bug")
• Decision analysis
• Evaluating arguments

When CoT adds little value:

• Simple factual lookups
• Creative writing
• Translation
• Summarization of clear text

Example:

Without CoT:

"A store buys apples for $0.40 and sells them for $0.60. If they sell 200 apples and have $30 in overhead costs, what is their profit?" [Model may give incorrect answer due to arithmetic shortcuts]

With CoT:

"A store buys apples for $0.40 and sells them for $0.60. If they sell 200 apples and have $30 in overhead costs, what is their profit? Think step by step." [Model calculates revenue, cost of goods, gross profit, then subtracts overhead — correct answer far more likely]

---

Technique 3: Few-Shot Examples

Showing the model examples of the format or style you want is more reliable than describing it. Especially useful for:

• Custom output formats
• Specific writing styles
• Classification tasks
• Extraction tasks with specific patterns

Example (extraction with few-shot):

Extract the product name, price, and availability from the following product descriptions. Format as JSON.

Example 1: Input: "The UltraBoost 22 running shoe is currently priced at $180 and is in stock." Output: {"product": "UltraBoost 22", "price": "$180", "availability": "in stock"}

Example 2: Input: "Our Premium Yoga Mat ($45) is temporarily out of stock." Output: {"product": "Premium Yoga Mat", "price": "$45", "availability": "out of stock"}

Now extract from: "The CloudFoam Sneaker retails for $95 and ships within 24 hours."

The model reliably follows the demonstrated pattern. Describing the desired format in words alone produces more variation.

How many examples? 1–3 examples is usually sufficient. More than 5 adds tokens without proportionally improving results for most tasks.

---

Technique 4: Structured Output Instructions

If you need consistent structured output (JSON, XML, tables, numbered lists), specify it explicitly. Modern LLMs (GPT-5, Claude 4) support JSON mode / structured output:

For casual use:

"Return your answer as JSON with keys 'summary', 'action_items', and 'priority'."

For API use: Use structured output parameters in the API (GPT-5's response_format: {type: "json_schema"}, Claude's tool use for structured extraction) to enforce the schema at the token level — not just in the instruction.

---

Technique 5: System Prompts

For any repeated or production use case, move your context and instructions into the system prompt rather than re-specifying them in every user message.

System prompt (set once):

"You are a customer service agent for TechStore. You help customers with order status, returns, and product questions. You have access to the customer's order history in the context. Always be concise and friendly. Never promise delivery dates you cannot confirm. If you cannot answer, offer to connect the customer with a human agent."

User message (varies):

"Where is my order #1234567?"

System prompts ensure consistent behavior across all conversations without repeating instructions.

---

Technique 6: Negative Instructions (What Not to Do)

Telling the model what to avoid is often as important as specifying what to include:

• "Do not include caveats about AI limitations"
• "Do not use bullet points — write in paragraphs"
• "Do not reference Wikipedia or generic sources — only cite peer-reviewed research"
• "Do not restate the question before answering"

Important: Negative instructions work reliably in GPT-5 and Claude 4. Earlier models sometimes did the opposite of negative instructions — if your model frequently violates "do not" instructions, rephrase as positive instructions instead ("respond concisely in 3 sentences or fewer").

---

Troubleshooting Bad Outputs

Problem	Likely Cause	Fix
Generic, vague response	Vague prompt	Add specific context, role, and format requirements
Model ignores instructions	Instructions buried in long prompt	Move critical instructions to the beginning and end
Hallucinated facts	Model fills knowledge gaps	Add "If you don't know, say so" or use RAG to provide facts
Wrong format	Format described but not demonstrated	Add few-shot examples
Too long / too short	No length constraint	Specify "in X sentences" or "in under X words"
Reasoning errors	No CoT	Add "Think step by step"
Inconsistent style	No style anchor	Provide a style example (few-shot) or a style guide reference

Claude 4 vs GPT-5: which model responds better to advanced prompting →