Design an Experiment or A/B Test
Turns a product hypothesis into a testable experiment with a defined metric, a sample-size reality check, and pre-committed decision rules — so the result gets interpreted honestly rather than argued about afterward. Use it before you ship the variant.
0 likes
0 dislikes
Sign in to rate this prompt
Prompt
You are an experimentation-literate product manager who cares more about learning something true than about winning.
What I want to test: {{idea}}
What I believe will happen and why: {{hypothesis}}
The metric I think it moves, and its current value: {{metric_and_baseline}}
Traffic or users available: {{traffic}}
How long I can run it: {{duration}}
What's at stake if it goes badly: {{risk}}
Produce:
**Sharpened hypothesis.** Restate {{hypothesis}} as a specific, falsifiable prediction: this change will move this metric by roughly this much, for this population, because of this mechanism. Name the mechanism — a test with no theory teaches nothing when it fails.
**Design.** Control and variant, what exactly differs, the unit of randomization, who's eligible, and what to do about users who never see the change.
**Metrics.** The primary decision metric, the secondary metrics that explain it, and the guardrails that stop the test early.
**Sample-size reality check.** Given {{traffic}}, {{metric_and_baseline}}, and {{duration}}, roughly what effect size is detectable. Show the reasoning. If the test is underpowered for a realistic effect, say so plainly and give me the alternatives: run longer, pick a more sensitive metric, test a bolder version, or skip the test and just decide.
**Decision rules, pre-committed.** What result means ship, what means kill, what means iterate, and what means inconclusive. Write these before the data exists — that's the entire point. Include the rule against peeking and stopping early on a good-looking day.
**Threats to validity.** Novelty effects, seasonality, contamination between groups, a concurrent launch, and any way {{idea}} could leak to the control group.
**Whether to test at all.** Given {{risk}} and {{traffic}}: is an experiment the right instrument? Sometimes the change is obviously right, the traffic is too low to ever resolve, or a qualitative study answers it faster. Say so if that's the case here.
**If it wins.** What we'd actually have learned, and what we'd still not know.