Engineering Judgment Drills

When performance plateaus or resource utilization doesn't match expectations, simply patching the code often masks the underlying problem. This skill implements a rigorous engineering judgment drill to help you understand exactly why your code behaves the way it does. Instead of providing blind fixes, it turns your project into a controlled instrument for measurement and learning.

What it does

• Benchmarks bottlenecks by creating standalone, non-invasive scripts to exercise specific code paths.
• Sweeps parameters across wide ranges to identify inflection points in wall time, CPU usage, and memory.
• Validates mental models by requiring verbatim predictions before any measurements are revealed.
• Stress-tests explanations through targeted attacks like concurrency shifts, data shape changes, and resource exhaustion.
• Audits reasoning via a structured rubric that reviews the precision of your technical judgment.

How it works

1. Scope and Confirm: The agent analyzes your code, picks a measurable question, and establishes a baseline run.
2. Predict and Measure: You provide predictions for specific axes while the agent runs sweeps using a dedicated Python tool.
3. Explain and Attack: You explain the resulting curves, and the agent attempts to invalidate that explanation with new experiments.
4. Review: The agent provides a final critique of your reasoning, identifying solid insights and superficial assumptions.

Frameworks & tools

This skill uses a bundled Python 3 utility for high-precision measurement (p50/p95 wall time, RSS memory, CPU usage). It requires no external network access and operates strictly on local processes and files.

Why this beats prompting it yourself

Generic prompts often lead to speculative answers or "hallucinated" performance advice based on general patterns rather than your specific hardware and runtime constraints. This skill enforces a "measurement first" gate, preventing any explanation that isn't backed by local CSV-based evidence.

Use cases

• Performance Plateaus: Determining why a system stops scaling past a specific number of threads or worker processes.
• Lock Contention: Testing how different synchronization mechanisms handle increased contention under high load.
• Resource Scaling: Quantifying exactly how memory consumption grows relative to input payload size.

Known limitations

Measurements are limited to the local machine and OS. Performance on clusters, GPUs, or production-only environments can only be simulated via provided commands, not executed directly.