Performancesmash

Perform a top-down performance audit. Focus on where time, memory, IO, and network are actually being wasted, and on which changes will materially improve latency, throughput, or resource efficiency.

Core Model

• Measure first. Guessing is how teams burn weeks for no gain.
• Treat data movement as a first-class cost: memory traffic, cache misses, serialization, network hops, disk reads, and cross-service chatter all count.
• Optimize the bottleneck that exists, not the one that is easiest to talk about.
• Prefer structural wins over heroics: fewer hops, fewer queries, less copying, better batching, better locality, better caching, simpler hot paths.
• Treat abstraction skeptically on hot paths. If indirection, layering, object churn, or generic machinery obscures the cost model, call it out.
• Broad before deep. First find whether the problem is CPU, memory, DB, network, lock contention, payload size, queueing, or coordination. Then drill into the dominant constraint.

Positioning

• Use this skill for repo-wide or subsystem-wide performance work.
• Use this skill when the user wants bottlenecks, optimization opportunities, or a diagnosis of latency, throughput, or resource waste before editing code.
• Prefer this over style-level cleanup when the real issue is performance shape rather than code tidiness.
• Be blunt. Do not get hypnotized by micro-optimizations when the real loss is query shape, chatty boundaries, unnecessary copies, or long dependency chains.