Evaluating Prompt Performance: From Vibes to Metrics

Early prompt tuning feels like cooking — taste, adjust, taste again. That works until you have three prompts in production serving 10,000 users. Then you need evals: a test set you run on every prompt change, with metrics that fail loudly when quality regresses.

Build an eval set

1. 1Collect 20-100 real inputs that represent your task (not synthetic fluff).
2. 2For each input, define what a correct or acceptable output looks like.
3. 3Include hard cases: ambiguous inputs, edge cases, adversarial inputs.
4. 4Version the set. Freeze it. Label updates v1, v2, v3.
5. 5Split into 'validation' (tune on) and 'held-out test' (final check).

Metric families

LLM-as-judge template

You are a rigorous evaluator. Score the following AI response.

<task_description>
{TASK}
</task_description>

<input>
{INPUT}
</input>

<response>
{MODEL_OUTPUT}
</response>

<reference_answer>
{IDEAL_ANSWER}
</reference_answer>

Score the response on:
1. Correctness (1-5): Does it answer the task accurately?
2. Completeness (1-5): Does it cover all parts of the task?
3. Format adherence (1-5): Does it match the required format?
4. Conciseness (1-5): Is it appropriately concise?

Respond in this XML format:
<scores>
  <correctness>N</correctness>
  <completeness>N</completeness>
  <format>N</format>
  <conciseness>N</conciseness>
</scores>
<justification>One paragraph explaining the scores.</justification>
<failure_mode>If any score is below 4, name the primary failure mode.</failure_mode>

The regression gate

Every proposed prompt change runs against the eval set automatically. If any metric drops by more than a threshold (e.g., -5% correctness), the change is blocked or flagged for human review. This is the same discipline as unit tests — just for prompts.

# Pseudocode for an eval run
for case in eval_set:
    output = run_prompt(candidate_prompt, case.input)
    scores = judge(task, case.input, output, case.ideal)
    results.append(scores)

mean = average(results)
if mean.correctness < baseline.correctness - 0.05:
    fail("Regression detected on correctness")
elif any_case.correctness < 3:
    warn("Individual case below threshold — review")
else:
    pass_and_record_new_baseline()

Beyond single-prompt eval

• Compare across models: same prompt on Claude Sonnet 4.5 vs Opus 4.7 vs Haiku.
• Compare across temperatures: find the sweet spot for your task.
• Compare prompt variants side-by-side (A/B testing at the prompt level).
• Track drift: run the same eval monthly; model updates can shift performance silently.
• Track cost/quality tradeoffs: the cheapest prompt that meets your quality bar often wins.