A/B Test Readout Skill

The hard part of an experiment is the readout: not "B won" but "is this real, is it big enough to matter, and should we ship?" This skill turns results into an honest decision — and flags the ways A/B results lie.

Working from a brief

Given results (even partial), write the full readout anyway. If significance isn't provided, reason about it from the numbers and flag what's needed to confirm. Mark assumed figures. Never declare a winner without addressing significance and sample.

Required Inputs

Ask for (if not already provided):

The hypothesis and the primary metric
Results — control vs variant: conversions/rate, sample size per arm, duration
Guardrail metrics (revenue, retention, latency, complaints) that mustn't regress
Pre-registered decision rule (what would count as a win) if one exists

Output Format

1. Verdict (one line)

Ship / Don't ship / Inconclusive — keep running — with the headline number.

2. The result

| Metric | Control | Variant | Relative lift | Significant? | |---|---|---|---|---| | Primary | | | | p / CI | | Guardrail(s) | | | | |

State statistical significance (p-value / confidence interval) and practical significance (is the lift big enough to matter given the cost?).

3. Did it really win?

Address the ways A/B tests mislead:

Sample / power — was the test adequately powered, or under-sampled?
Peeking — was the call made early, inflating false positives?
Novelty / primacy — could the effect fade?
Segments — does the win hold across key segments, or is it driven by one?

4. Segment cuts

Where the effect is strong vs flat vs negative (new vs returning, platform, geography).

5. Recommendation & next step

Ship / iterate / re-run, plus what to monitor post-launch or what the follow-up test should isolate.

Quality Checks

[ ] Distinguishes statistical from practical significance
[ ] Checks guardrail metrics, not just the primary
[ ] Flags peeking, power, novelty, and segment-driven wins
[ ] Recommendation follows from the evidence, with a monitoring/next-test step
[ ] Doesn't declare a winner on an underpowered or peeked result

Anti-Patterns

"B won by 8%!" with no significance or sample size
Calling a result early (peeking) and shipping
Ignoring a guardrail regression because the primary went up
A statistically significant but practically meaningless lift treated as a win