prompt-testing · diff
git:20260705.ba8ee52 to git:20260729.3b91eed
7 added, 1 removed. Audit A to A.
---
name: prompt-testing
- description: "A/B testing and performance metrics for prompts. Use when: comparing two prompt variants, defining quality/efficiency/robustness metrics, or deciding whether to adopt a challenger prompt over a baseline."
+ description: "Use when comparing two prompt variants, defining quality/efficiency/robustness metrics, or deciding whether to adopt a challenger prompt over a baseline."
allowed-tools: Read, Write, Bash
---
+
+ <objective>
+ Prompt Testing runs A/B comparisons between prompt variants through a 5-step workflow: define the objective and metrics, prepare variants A/B and a test dataset, execute on the dataset, analyze and compare results, then decide. Metrics span three categories -- quality (accuracy, compliance, consistency, relevance), efficiency (input/output tokens, latency, cost), and robustness (edge-case handling, jailbreak resistance, error recovery) -- plus a UX category detailed in `metrics.md`.
+
+ The adoption decision is rule-based: adopt B if its accuracy is at least equal with acceptable token cost, consider B as a trade-off if accuracy improves >10% despite <20% token regression, otherwise keep A or iterate. Requires a minimum of 20 test cases with 15-20% edge cases for statistical significance.
+ </objective>
# Prompt Testing
Skill for testing, comparing, and measuring prompt performance.
## References
- [metrics.md](references/metrics.md) - Load when: defining or scoring Quality/Efficiency/Robustness/UX metrics with thresholds and calculation formulas
- [methodology.md](references/methodology.md) - Load when: running a full A/B test (hypothesis, dataset sizing, statistical significance, common pitfalls)
- [templates.md](references/templates.md) - Load when: writing a test dataset JSON or an A/B test report
## Testing Workflow
```text
1. DEFINE
└── Test objective
└── Metrics to measure
└── Success criteria
2. PREPARE
└── Variants A and B
└── Test dataset
└── Baseline (if existing)
3. EXECUTE
└── Run on dataset
└── Collect results
└── Document observations
4. ANALYZE
└── Calculate metrics
└── Compare variants
└── Identify patterns
5. DECIDE
└── Recommendation
└── Statistical confidence
└── Next iterations
```
## Performance Metrics
### Quality
| Metric | Description | Calculation |
|--------|-------------|-------------|
| **Accuracy** | Correct responses | Correct / Total |
| **Compliance** | Format adherence | Compliant / Total |
| **Consistency** | Response stability | 1 - Variance |
| **Relevance** | Meeting the need | Average score (1-5) |
### Efficiency
| Metric | Description | Calculation |
|--------|-------------|-------------|
| **Tokens Input** | Prompt size | Token count |
| **Tokens Output** | Response size | Token count |
| **Latency** | Response time | ms |
| **Cost** | Price per request | Tokens × Price |
### Robustness
| Metric | Description | Calculation |
|--------|-------------|-------------|
| **Edge Cases** | Edge case handling | Passed / Total |
| **Jailbreak Resist** | Bypass resistance | Blocked / Attempts |
| **Error Recovery** | Error recovery | Recovered / Errors |
For full definitions, thresholds, and the UX metrics category, see [metrics.md](references/metrics.md). For the test dataset and report formats, see [templates.md](references/templates.md).
## Commands
```bash
# Create a test
/prompt test create --name "Test v1" --dataset tests.json
# Run an A/B test
/prompt test run --a prompt_a.md --b prompt_b.md --dataset tests.json
# View results
/prompt test results --id test_001
# Compare two tests
/prompt test compare --tests test_001,test_002
```
## Decision Criteria
### When to adopt variant B?
```text
IF:
- Accuracy B >= Accuracy A
AND (Tokens B <= Tokens A * 1.1 OR accuracy improvement > 5%)
AND no regression on edge cases
THEN:
→ Adopt B
ELSE IF:
- Accuracy improvement > 10%
AND token regression < 20%
THEN:
→ Consider B (acceptable trade-off)
ELSE:
→ Keep A or iterate
```
## Best Practices
1. **Minimum 20 test cases** for significance
2. **Include edge cases** (15-20% of dataset)
3. **Test multiple runs** for consistency
4. **Document hypotheses** before testing
5. **Version the prompts** being tested