Use this skill to reason about bottlenecks before the project burns time on blind optimization.
Core Workflow
- State the workload, SLOs, hardware targets, and failure envelope.
- Draw the critical path and concurrency model.
- Build a budget table for latency, bandwidth, memory, storage, and power.
- Identify queueing, serialization, synchronization, and data-movement hotspots.
- Propose measurement points before proposing optimizations.
- Rank fixes by expected impact, risk, and implementation cost.
Execution Rules
- Do not optimize without a budget.
- Separate steady-state behavior from tail behavior.
- Check batching, caching, affinity, backpressure, retries, and failure recovery.
- For accelerators and edge systems, include host-device transfer and power or thermal constraints.
- Be explicit about whether a claim is analytical, measured, or inferred.
Output Contract
Return: