Advanced Prompt Evaluation

Why Evaluation Matters

Most teams tune prompts by feel — tweak, test manually with 3 examples, ship. This works until it doesn't: a prompt change that improves one case breaks five others. Evaluation (evals) catches regressions, measures improvement, and gives you confidence to iterate fast.

Building an Eval Dataset

An eval dataset is a list of (input, expected_output) pairs. Start small — 20 good examples beats 200 mediocre ones.

# evals/dataset.json
[
  {
    "id": "triage_001",
    "input": "My invoice from last month is wrong — I was charged twice",
    "expected": {"category": "billing", "priority": "high"},
    "tags": ["billing", "duplicate_charge"]
  },
  {
    "id": "triage_002",
    "input": "How do I export my data to CSV?",
    "expected": {"category": "technical", "priority": "low"},
    "tags": ["technical", "data_export"]
  },
  {
    "id": "triage_003",
    "input": "Your app crashes every time I open it on iOS 17",
    "expected": {"category": "technical", "priority": "high"},
    "tags": ["technical", "crash", "mobile"]
  }
  // ... 17 more examples covering edge cases
]

Running Evals: Exact Match + LLM-as-Judge

import json, anthropic
client = anthropic.Anthropic()

def run_prompt(user_input: str) -> dict:
    """The function under test"""
    r = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=128,
        system="""Classify the support ticket. Return JSON only:
{"category": "billing|technical|feature_request|other",
 "priority": "high|medium|low"}""",
        messages=[{"role": "user", "content": user_input}]
    )
    return json.loads(r.content[0].text)

def exact_match_eval(dataset: list) -> dict:
    """For structured outputs with known ground truth"""
    results = {"total": len(dataset), "correct": 0, "failures": []}
    for example in dataset:
        predicted = run_prompt(example["input"])
        if (predicted["category"] == example["expected"]["category"] and
            predicted["priority"] == example["expected"]["priority"]):
            results["correct"] += 1
        else:
            results["failures"].append({
                "id": example["id"],
                "input": example["input"],
                "expected": example["expected"],
                "got": predicted
            })
    results["accuracy"] = results["correct"] / results["total"]
    return results

LLM-as-Judge (for Open-Ended Outputs)

When outputs are prose (summaries, answers, emails), there's no single correct answer. Use Claude itself as the judge — it's surprisingly reliable with a well-structured rubric.

def llm_judge(question: str, ideal_answer: str,
              actual_answer: str) -> dict:
    """Score actual_answer against ideal on a 1-5 scale."""
    prompt = f"""You are an expert evaluator. Score the ACTUAL answer
against the IDEAL answer on these dimensions (1-5 each):
- Accuracy: Does it contain correct information?
- Completeness: Does it cover the key points?
- Clarity: Is it clear and well-written?
- Conciseness: Does it avoid unnecessary filler?

Question: {question}
Ideal answer: {ideal_answer}
Actual answer: {actual_answer}

Return JSON only:
{{"accuracy": N, "completeness": N, "clarity": N,
  "conciseness": N, "overall": N, "reasoning": "one sentence"}}"""

    r = client.messages.create(
        model="claude-opus-4-5",   # use a strong model as judge
        max_tokens=256,
        messages=[{"role": "user", "content": prompt}]
    )
    return json.loads(r.content[0].text)

# Score an entire dataset
scores = [
    llm_judge(ex["question"], ex["ideal"], run_prompt(ex["question"]))
    for ex in open_ended_dataset
]
avg_overall = sum(s["overall"] for s in scores) / len(scores)
print(f"Average score: {avg_overall:.2f}/5")

Prompt Versioning & CI

# prompts/v2_support_triage.py
SYSTEM_PROMPT = """..."""  # version controlled in Git

# Run evals on every PR
# .github/workflows/evals.yml
# - python -m pytest evals/test_support_triage.py --threshold=0.85

import pytest
from evals.runner import exact_match_eval
from evals.dataset import load_dataset

def test_support_triage_accuracy():
    dataset = load_dataset("support_triage")
    results = exact_match_eval(dataset)
    assert results["accuracy"] >= 0.85, (
        f"Accuracy {results['accuracy']:.0%} below 85% threshold. "
        f"Failures: {results['failures']}"
    )

Hands-on: Build Your First Eval Suite