Overview
Premature optimization is the root of all evil. But ignoring performance until it's a crisis is equally harmful. This skill enforces data-driven optimization: profile first, optimize the bottleneck, measure the improvement.
When to Use
- When performance issues are reported in production
- Before optimizing any code (to ensure you're optimizing the right thing)
- When reviewing changes that touch performance-sensitive paths
Process
Step 1: Measure the Baseline
- Reproduce the performance issue reliably.
- Measure current performance: latency p50/p95/p99, throughput, memory, CPU.
- Profile to find the actual bottleneck — not where you think it is.
- Write the performance test you'll use to validate improvement.
Verify: You have concrete baseline numbers, not gut feelings.
Step 2: Identify the Real Bottleneck
- Use profiling tools: flame graphs, CPU profiles, memory profiles.
- Find the top 3 hotspots by actual execution time (not lines of code).
- The bottleneck is rarely where you expect it to be. Trust the data.
Verify: Bottleneck identified by profiling data, not assumption.
Step 3: Optimize Only the Bottleneck
- Fix only the profiled bottleneck — nothing else.
- Common optimizations by type:
- CPU: Algorithmic improvement (O(n²) → O(n log n)), caching, batching
- Memory: Streaming instead of buffering, object pooling, lazy loading
- I/O: Connection pooling, N+1 query elimination, caching, async/parallel calls
- AI: Prompt caching, batch inference, smaller models for simpler tasks
Verify: Change targets the profiled bottleneck, not speculative improvements.
Step 4: Measure the Improvement
- Run the same performance test from Step 1.
- Compare before vs. after metrics.
- If improvement < 20%: the optimization may not be worth the complexity.
Verify: Improvement measured with the same test harness as baseline.
Common Rationalizations (and Rebuttals)
| Excuse | Rebuttal |
|---|---|
| "I know where the bottleneck is" | You're probably wrong. Profile first. |
| "This is clearly slow" | "Clearly slow" rarely matches profiler output. Measure. |
| "We'll optimize later" | If it's slow enough to mention, it's slow enough to measure now. |
Verification
- Baseline metrics captured before any optimization
- Bottleneck identified by profiler (not assumption)
- Optimization targets only the profiled bottleneck
- Improvement measured with same test harness
- Before/after numbers documented