Shihipar's Claude Code workflow: build on low, verify on high

Claude Code's effort setting mostly buys extra verification, not a smarter plan, per Shihipar's workflow for spending effort. That matters for token budgets. On Terminal-Bench 3.0, the same task took Fable 5.1 a median 73k tokens at low effort and 222k at max, a 3x jump.

The extra tokens buy real accuracy. Security task accuracy was 64% at low effort and 87% at high. Hardware task accuracy was 34% at low and 75% at high.

The workflow starts with an interview against a spec. Shihipar has Claude implement and iterate on low effort, then switches to high effort for verification and tests once the code exists.

If a task still fails at high effort, the fix is a better spec. Turning the dial to max just spends more tokens hunting for an approach that was never going to work.

The post doesn't say how this holds outside Terminal-Bench 3.0, or whether it applies to tasks that aren't security or hardware focused.