Prompt Engineering / Prompt Optimization
Optimize a prompt for a specific metric (accuracy, brevity, creativity, safety) with trade-off analysis.
Why This Prompt Exists
“Good prompt” is not a metric. Different tasks need different trade-offs — but most prompts optimize for nothing specific.
You get:
- prompts that are accurate but verbose (when you need brevity)
- prompts that are creative but inconsistent (when you need reliability)
- prompts that are safe but unhelpful (overly cautious)
- no ability to tune for your specific use case
- trade-offs that are implicit, not explicit
But optimization requires metrics:
- accuracy: correct vs. incorrect (binary or scored)
- brevity: output length, token count
- creativity: novelty, diversity, unexpectedness
- safety: harmful output rate, refusal rate
- consistency: same output for same input
- latency: time to first token
- cost: total tokens (input + output)
Without metrics, you optimize for “feels better” — which is unreliable.
This prompt optimizes prompts for specific, measurable metrics.
The Prompt
Assume the role of a prompt optimization engineer who optimizes for measurable metrics. Your task is to modify a prompt to improve a target metric. Generate: 1. BASELINE PERFORMANCE - Current prompt - Target metric: [name] = [value] (from testing) - Other metrics (for trade-off awareness) 2. OPTIMIZATION TARGET - Metric to improve: [accuracy / brevity / creativity / safety / consistency / latency / cost] - Current value: [X] - Target value: [Y] - Acceptable degradation in other metrics: [list with limits] 3. PROMPT MODIFICATIONS (test these in order) | Modification | Expected effect on target metric | Expected effect on other metrics | Rationale | |--------------|--------------------------------|--------------------------------|-----------| | [add instruction] | +[X]% | [e.g., may increase tokens] | [why it works] | | [add example] | +[X]% | [e.g., minimal effect] | [why it works] | | [remove instruction] | +[X]% | [e.g., may reduce accuracy] | [why it works] | 4. PROPOSED OPTIMIZED PROMPT - Full prompt with recommended modifications 5. TRADE-OFF ANALYSIS - What you gain (improvement in target metric) - What you lose (degradation in other metrics) - Is the trade-off worth it? (Yes/No/Maybe — test to find out) 6. TESTING RECOMMENDATIONS - How to validate improvement (A/B test, sample size) - What to monitor after deployment INPUTS: Current prompt: [PASTE THE PROMPT] Target metric and baseline (from testing): [E.G., "Accuracy = 85% on test set of 200 examples"] Other metrics and baseline (optional): [E.G., "Average output length = 150 tokens"] Acceptable trade-offs: [E.G., "Can increase output length up to 200 tokens, can reduce accuracy no more than 2%"] Task type: [CLASSIFICATION / GENERATION / EXTRACTION / OTHER] RULES: - Measure before you optimize (you need a baseline) - Optimize one metric at a time (multi-objective optimization is complex) - Trade-offs are inevitable — be explicit about what you're sacrificing - Test proposed changes (don't assume they'll work) - If you can't measure it, you can't optimize it
How To Use It
- Measure baseline performance before optimizing — you need a starting point.
- Optimize one metric at a time — trying to improve everything usually improves nothing.
- Be explicit about acceptable trade-offs — “accuracy can drop 2% for 50% shorter outputs”
- Test each modification separately before combining them.
- Re-measure after each change to ensure you’re actually improving.
Example Input
Current prompt:
“Summarize this article in a few sentences.”
Target metric and baseline:
“Brevity = average 120 tokens per summary (measured on 100 articles)”
Other metrics and baseline:
“Accuracy (factual correctness) = 90% on human evaluation”
Acceptable trade-offs:
“Can reduce accuracy up to 5% (to 85%) if brevity improves significantly (to under 60 tokens)”
Why It Works
Most prompt optimization is subjective — “this version seems better” — which is impossible to reproduce or trust.
This framework improves outcomes by forcing:
- baseline measurement (where are we starting?)
- explicit optimization target (what metric are we improving?)
- trade-off specification (what are we willing to lose?)
- testable modifications (not just “try things”)
- validation recommendations (how to confirm improvement)
Great metric-based optimization doesn’t guess — it measures, changes, and measures again.
Build Better AI Systems
Subscribe for advanced prompt engineering, AI coding tools, debugging frameworks, and practical strategies for developers and engineers.