Verification

One at a time. Tell it whether you knew, and the ones you did not stay in your pile.

1 / 7

Ask Claude for the command it ran and its output.

Evidence beats "it should work". Give it a pass or fail check: a test file, a build exit code, a linter, a fixture diff or a screenshot.

The whole topic7

Ask Claude for the command it ran and its output.
Evidence beats "it should work". Give it a pass or fail check: a test file, a build exit code, a linter, a fixture diff or a screenshot.
/verify builds, runs and looks at the app.
It catches what tests and typechecks miss.
/code-review uses a fresh subagent that never saw your conversation.
So it is not anchored on your reasoning. Add --fix to apply the findings.
/simplify runs four review agents in parallel on your changed code.
Run it before you commit a large change.
The agent solves whatever the verifier measures, so the verifier must be nearly perfect.
In the C compiler project, GCC served as a known-good oracle, which let agents isolate bugs in parallel.
pass@k rises with k, but pass^k falls: 75 percent per trial gives about 42 percent for three in a row.
If an agent must succeed every time, measure consistency, not best of many.
A 0 percent pass rate over many trials usually means a broken task, not a weak agent.
Check the task and the grader before you blame the model.