Improving Accuracy

Introduction

Beyond hallucination reduction, accuracy improvements come from: better examples, better format, CoT, self-consistency, and the right model for the task.

Beginner analogy: Like tuning a guitar — many small knobs, not one big lever.

Understanding the topic

Core concepts to understand:

• Add 2–3 diverse few-shot examples.
• Use CoT for reasoning tasks.
• Self-consistency for high-stakes single answers.
• Switch model only after prompt is exhausted.
• Specialise prompt per intent (route → expert prompt).

Syntax reference

Visual workflow / architecture:

Accuracy levers (in order)
1. Better instructions
2. Better format
3. Few-shot examples
4. CoT
5. Self-consistency
6. Better model

Real-world use

Most accuracy wins in production come from prompt levers, not model swaps. The same Claude 3.5 can swing 30 points based on prompt quality.

Best practices

• Apply cheap levers (instructions, format) before expensive ones (more examples, bigger model).
• Specialise: classify intent, then use the right expert prompt.
• Measure improvement on the eval set, not anecdotes.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Order accuracy levers cheap → expensive.
• Q2. Why specialise prompts per intent?
• Q3. When swap models?
• Q4. How prove a change improved accuracy?