| name | growth-experiments |
| description | Designs statistically sound A/B experiments and interprets results: defines hypothesis and primary metric, calculates required sample sizes, flags confounds, and—given results—applies the appropriate significance test and states a clear conclusion. Use this skill when the user says "design an A/B test", "is this result significant", "how long should I run this experiment", "analyze my experiment results", "help me set up a test", or wants to avoid underpowered or inconclusive tests. Not for multi-armed bandits, qualitative usability research, or instrumentation and tracking implementation. |
| license | MIT |
| metadata | {"skillsmith-invocation":"user","skillsmith-maturity":"experimental"} |
| user-invocable | true |
growth-experiments
An underpowered experiment wastes traffic and produces nothing; a
mislabeled winner produces a confident regression. This skill designs
tests that are sized to detect the effect that would matter and
interprets results without overclaiming.
Determine mode
Two modes; detect from context:
- Design — user wants to plan a new experiment. Requires:
what is being tested, the primary metric, and a baseline value for
that metric. Clarify if missing.
- Analyze — user has results. Requires: sample sizes per
variant, conversions or metric values per variant, and the original
hypothesis (what direction of change was expected). Clarify if
missing.
Design mode
Write the hypothesis first. Form: "If [change], then [metric]
will [direction] because [mechanism]." A hypothesis without a
mechanism is a guess; a mechanism makes the result interpretable
regardless of outcome.
State the primary metric and guardrails. One primary metric per
test — the one that would change the decision. Guardrail metrics are
those that must not degrade (e.g., downstream conversion,
unsubscribe rate). Name them explicitly; failing a guardrail
overrides a primary metric win.
Calculate minimum detectable effect (MDE) and sample size. Ask
the user: what is the smallest improvement that would change a
decision? Use that as MDE. Calculate required sample size per
variant at 80% power and 95% confidence (two-tailed). State how
long the test needs to run given current traffic, and flag if the
required runtime exceeds 4 weeks — long tests accumulate time-
series confounds.
Flag pre-experiment confounds. Check for: novelty effect
(change that wears off), seasonality (test spans a known traffic
shift), leakage (variants visible to the same user via different
paths), and multiple comparisons (more than 2 variants without
correction).
Analyze mode
Apply the appropriate test:
- Binomial outcome (clicks, conversions): two-proportion z-test.
- Continuous outcome (revenue, time on page): Mann-Whitney U (non-
parametric; does not assume normality for skewed distributions).
Report: observed effect size, p-value, 95% confidence interval on
the effect, and a plain-language conclusion. Do not use
"statistically significant" without also stating what the effect
means in absolute and relative terms.
Flag if the test was stopped early — peeking invalidates the p-
value. If it was stopped early, state the result is directional,
not conclusive, and recommend a pre-registered stopping rule for the
next test.
Flag guardrail metrics before declaring a winner. A primary metric
win with a guardrail regression is a loss.