How Do You Evaluate Whether a Prompt Is Working Well?

Why This Is Asked

Prompt evaluation is a fundamental skill for anyone building LLM-powered features. Interviewers ask this to see if you approach prompt development systematically (like an engineer) or intuitively (like a hobbyist). Strong candidates treat prompts like code: they need tests, version control, and regression tracking.

Key Concepts to Cover

• Golden dataset — curated input/expected output pairs for evaluation
• Automated metrics — when exact match, ROUGE, or LLM-as-judge are appropriate
• Human evaluation — what requires human judgment and how to structure it
• Regression testing — ensuring prompt changes do not break existing behavior
• A/B testing — comparing prompt versions in production
• Coverage — ensuring eval set covers edge cases, not just happy paths

How to Approach This

1. Define What "Working" Means

Before evaluating, define success:

• What is the task? (classification, extraction, generation, QA?)
• What does a correct output look like?
• What failure modes matter most?

2. Build a Golden Evaluation Dataset

A golden dataset is a set of (input, expected output) pairs:

• Start with 20-50 examples covering common cases
• Include edge cases: empty inputs, unusual formats, adversarial inputs
• Include known failure cases from production
• Label expected outputs carefully

3. Choose the Right Metrics

| Task Type | Good Metrics | |-----------|-------------| | Classification | Accuracy, F1, precision/recall | | Structured extraction | Field-level accuracy, JSON validity | | Summarization | ROUGE, human rating | | Open-ended generation | LLM-as-judge, human preference | | Q&A | Exact match, semantic similarity |

LLM-as-judge: Use a stronger LLM to evaluate outputs. Define a rubric (e.g., "Rate this answer 1-5 for accuracy, helpfulness, conciseness"). Cost-effective and scalable.

4. Run Evaluation Systematically

for example in golden_dataset:
    output = call_llm(prompt_template, example.input)
    score = evaluate(output, example.expected)
    results.append(score)

5. Regression Testing

Any time you change a prompt, run the full eval set and compare scores. Track metrics over time to catch gradual degradation.

Common Follow-ups

1. "What if you don't have ground truth labels?" Use LLM-as-judge, collect labels from user feedback, start small with human-labeled set.

2. "How do you evaluate prompts for subjective tasks like tone or style?" Preference-based evaluation (A vs. B), rubric-based human rating, embedding-based similarity.

3. "How do you handle evaluation when the same input can have multiple valid outputs?" Reference-free evaluation using LLM-as-judge, checking underlying intent rather than string matching.