Prompt Engineering for Developers: Beyond Basic Chat

Prompts Are Code

Treat your prompts like code: version them, test them, and review them. A poorly written prompt will give inconsistent results no matter how powerful the model.

Technique 1: Structured Output

Always request structured output for programmatic use:

Analyze the following code review comment and return JSON:
{
  "severity": "critical" | "warning" | "suggestion",
  "category": "bug" | "style" | "performance" | "security",
  "summary": "one line summary",
  "suggestion": "proposed fix"
}

Technique 2: Chain of Thought

For complex reasoning, make the model show its work:

Think step by step: 1. Identify the tables and joins 2. Check for missing indexes 3. Look for N+1 query patterns 4. Estimate the query complexity

Then provide your recommendations. ```

Technique 3: Few-Shot Learning

Show examples of desired input/output pairs:

Example 1: User story: "As a user, I want to reset my password" Tasks: - Add /forgot-password route - Create email template - Implement token generation

Now convert: User story: "As a user, I want to export my data as CSV" ```

Evaluation

Build an eval suite for your prompts:

const testCases = [
  { input: "...", expected: "...", metric: "exact_match" },
  { input: "...", expected: "...", metric: "semantic_similarity" },
]

Run evals on every prompt change, just like unit tests for code.