Prompt Engineering

Core Prompt Patterns

Prompt patterns are reusable strategies for structuring your instructions to an LLM. Choosing the right pattern is the single highest-leverage decision you make — it determines whether the model reasons carefully or guesses wildly.

Zero-Shot vs Few-Shot Decision

The most common mistake in prompt engineering is providing too many examples when none are needed, or providing none when the model clearly needs guidance. This decision tree helps you choose the right approach for your task.

Is the task straightforward and well-defined?
  YES -> Zero-shot (add "Think step by step" if reasoning needed)
  NO  -> Does the output need a specific format?
           YES -> Few-shot (2-3 examples showing format)
           NO  -> Does it involve nuanced judgment?
                    YES -> Few-shot with edge cases
                    NO  -> One-shot with clear instructions

Chain-of-Thought Templates

Chain-of-thought prompting forces the model to show its reasoning before giving an answer. This dramatically improves accuracy on math, logic, and multi-step reasoning tasks — often turning a wrong answer into a correct one.

Basic CoT

Q: {question}
A: Let's think step by step.

Structured CoT

Solve the following problem. Show your reasoning in these steps:
1. Identify the key information
2. Determine the approach
3. Execute step by step
4. Verify the answer

Problem: {problem}

Self-Consistency (with CoT)

Generate 5 independent solutions to this problem.
For each, think step by step.
Then compare all answers and select the most common result.

System Prompt Architecture

The system prompt is your contract with the model — it defines who it is, what it does, and how it behaves. A well-structured system prompt eliminates 80% of output quality issues before they happen.

Anatomy of an Effective System Prompt

You are {role} with expertise in {domain}.

## Task
{What to do}

## Constraints
- {Constraint 1}
- {Constraint 2}

## Output Format
{Exact format specification}

## Examples
Input: {example_input}
Output: {example_output}

System Prompt Checklist

Temperature and Sampling Parameters

Temperature and sampling parameters control how the model selects the next token. Getting these wrong means either robotic, repetitive outputs (too low) or hallucinated, incoherent ones (too high). Most production issues trace back to incorrect parameter settings.

Recommended Settings by Task

Rule of thumb: Set temperature OR top_p, not both. Use temperature for most cases.

Prompt Templates by Use Case

These are battle-tested templates for the most common LLM tasks. Copy, paste, and customize — they handle the structural patterns so you can focus on your specific domain requirements.

Classification

Classify the following text into exactly one category:
Categories: {list}

Text: "{input}"

Respond with only the category name, nothing else.

Extraction

Extract the following fields from the text below.
Return as JSON. Use null for missing fields.

Fields: {field_list}

Text: "{input}"

Summarization

Summarize the following text in {N} bullet points.
Each bullet should be one concise sentence.
Focus on: {key_aspects}

Text: "{input}"

Code Generation

Write a {language} function that {description}.

Requirements:
- {req1}
- {req2}

Input: {input_type}
Output: {output_type}

Include error handling and type hints.

Output Formatting Tricks

Unstructured LLM output is the #1 source of downstream bugs in production systems. These formatting techniques ensure your model returns parseable, consistent output every time.

Advanced Techniques

When a single prompt can't handle the full complexity of your task, these techniques let you decompose, delegate, and iterate. Prompt chaining alone can turn a 60% accuracy task into a 95% one.

Prompt Chaining

Step 1: Extract key entities -> {entities}
Step 2: For each entity, gather context -> {context}
Step 3: Synthesize into final answer using {entities} + {context}

Meta-Prompting

You are a prompt engineer. Write the optimal prompt for an LLM
to accomplish this task: {task_description}

The prompt should include:
- Clear role definition
- Step-by-step instructions
- Output format specification
- 2-3 examples

Negative Prompting

Do NOT:
- Include disclaimers or warnings
- Use bullet points
- Exceed 100 words
- Mention that you're an AI

Token Optimization

Every token costs money and adds latency. In production systems handling millions of requests, a 30% reduction in prompt tokens can save thousands of dollars per month without sacrificing quality.

Common Pitfalls

These are the mistakes every practitioner makes at least once. Bookmark this table — when your LLM output looks wrong, the fix is almost always here.

Evaluation Quick Checks

Before shipping any prompt to production, run these five checks. A prompt that works once in a playground can fail spectacularly at scale — these checks catch the gaps.

Quick Reference: Prompt Length Guidelines

Context windows are not infinite — and models don't attend equally to all content. Understanding your token budget prevents truncated outputs and ensures the model actually reads your most important instructions.

Remember: Longer context windows do not mean the model attends equally to all content. Place critical instructions at the beginning and end of the prompt.