Validation & Model Selection

Teach Claude how to verify its own work — build commands, test runners, debug logs, emulator control, and more.

Validation loops are the single most important concept. Examples: Have Claude add debug logs and read them via tail. Use Puppeteer/Chrome to navigate and test web UIs. Hook into Perfetto for performance traces. Write integration tests. The more verification Claude can do autonomously, the better its output.

Use Opus for complex reasoning tasks. The slower speed is offset by running multiple Claude instances in parallel.

Opus produces better results for complex work. The author runs multiple Claude instances simultaneously, so the slower speed of Opus doesn't bottleneck productivity — while one instance processes, you work on another.

Watch Claude's thinking output for red flags like "I'm not sure" or assumptions. Interrupt and correct early.

Claude's extended thinking reveals its reasoning process. Look for uncertainty markers, incorrect assumptions, or wrong directions. Interrupt immediately when you spot issues — early course correction is far cheaper than fixing bad output.