Try it and inspect the result

Run the smallest safe experiment that answers the actual question. Use existing harnesses and real project inputs where practical. A scratch probe is enough when a permanent test would add no value.

Choose an observable outcome: actual API shape, elapsed time, persisted state, rendered appearance, or a reproducible failure. Inspect surprising results for harness mistakes before building on them. Distinguish observation from its explanation, then test the proposed cause.

Compare a design direction

For an aesthetic choice, produce a few meaningful alternatives and compare them at the intended scale. Keep presentation comparable. Judge subject fit, hierarchy, legibility, and the desired experience. A fixed variant count or number of rounds is unnecessary.

Research an unfamiliar technique when it would change the experiment; a reversible prototype can proceed alongside research. A specimen board supports a design decision. Integrate the chosen result into the actual product before claiming the requested implementation is complete.

Respect the evidence boundary

A still cannot establish smooth motion, and offline sound metrics cannot establish live playback or pleasantness. For temporal defects, exercise normal, rapid, and interrupted paths and inspect state ordering. A non-reproduction does not invalidate a user's repeatable observation.

Keep the setup proportional to the question. Think through costly or hard-to-reverse actions before executing them. Stop experimenting when the decision is supported and continue the requested implementation or verification. Report what ran, what was observed, and any unresolved limitation.