performance

Measure first. Optimize second. Never guess where the bottleneck is. Premature optimization is the root of all evil—but so is ignoring real problems. Performance is a feature. Latency is a bug. Have benchmarks or profiling data before making performance changes. If no data exists, create minimal reproduction to measure. Skill-specific: skills/performance/reference/performance-research.md

<core_principles>

1. MEASURE FIRST: Profile before optimizing. Data beats intuition.
2. IDENTIFY BOTTLENECK: Fix the slowest part first. Amdahl's Law applies.
3. ONE CHANGE AT A TIME: Measure after each change. Verify improvement.
4. REGRESSION TESTS: Performance tests prevent future degradation.
5. GOOD ENOUGH: Stop when requirements are met. Don't gold-plate. </core_principles>

<profiling_checklist> Before optimizing:

• Have reproduction case
• Baseline measurement documented
• Target performance defined
• Profiler identified hotspots

After optimizing:

• New measurement taken
• Improvement verified
• No functionality regression
• Performance test added </profiling_checklist>

<common_bottlenecks>

• N+1 queries (fetch in loop instead of batch)
• Missing indexes on frequently queried columns
• Synchronous I/O in hot paths
• Excessive serialization/deserialization
• Memory allocations in tight loops
• Blocking operations in async code </common_bottlenecks>

<anti_patterns>

• Optimizing without profiling data
• Micro-optimizing code that runs once
• Caching everything (cache invalidation is hard)
• Async for CPU-bound work
• Premature parallelization </anti_patterns>

<on_complete> agentdb write-end '{"skill":"performance","bottleneck":"X","measured_before":true,"improvement":"Xms|none"}' </on_complete>