Task Success Rate field

Average Corrections per Task field

Tool Call Efficiency field

User Satisfaction Score field

Response Latency field

Token Efficiency Ratio field

Instruction misunderstanding category

Output format errors category

Context loss category

Tool misuse category

Constraint violations category

Edge case handling category

Correction patterns analysis

Positive feedback patterns

Six test categories

Hallucination metric

30-day analysis period

Chain-of-thought step phrase

Self-verification checkpoint phrase

Good example Input field

Good example Reasoning field

Good example 'Why this works' field

Bad example 'Why this fails' field

Bad example 'Correct approach' field

Role: core purpose

Role: constraints section

Role: success criteria

Constitutional principle: factual accuracy

Constitutional principle: format validation

Constitutional principle: consistency

Critique-and-revise loop

Simple-to-complex ordering

Version format compliance

MINOR for prompt improvements

Git-based prompt storage

Alpha stage at 5%

Beta stage at 20%

Canary progression to 50% then 100%

7-day monitoring window

Rollback trigger: success rate drop >10%

Rollback trigger: critical errors >5%

Rollback trigger: cost increase >20%

Rollback process: 5 steps

Success: 15% improvement threshold

Success: 25% corrections reduction

Success: cost within 5%