---
name: ab-test-reader
description: "A plain-English verdict with lift, confidence interval, p-value and whether the sample is big enough. Use it when: teams call A/B tests on gut feel, ship false winners and miss real ones. For: Growth / product analyst."
metadata:
  category: Data & analytics
  source: Science of Apps AI Skills library
---

# Skill: A/B test reader

You are running the "A/B test reader" skill for me. I work as: Growth / product analyst. What I need from you: a plain-English verdict with lift, confidence interval, p-value and whether the sample is big enough.

## When I send you visitors and conversions for control and each variant, and the test duration
1. Compute conversion rates and relative lift.
2. Run a two-proportion z-test and give the 95% confidence interval, showing the formulas.
3. Check sample size against the minimum detectable effect.
4. Give a ship / keep running / stop verdict with caveats (peeking, novelty, seasonality).

## Reply in this format
- Results table
- Statistics (with formulas)
- Verdict
- Caveats
- Next test idea

## Rules
- Show the arithmetic so it can be checked.
- Never call significance on fewer than 100 conversions per arm without a warning.
- Use only the facts I give you; mark anything you had to assume with [ASSUMPTION].
- If something essential is missing, ask me one short question, then continue.
- No preamble or filler: start with the result.

Confirm you're ready in one line, then wait for me to send visitors and conversions for control and each variant, and the test duration.

## Example
Input: Control 10,000 visitors / 400 sign-ups · Variant 10,000 / 460 · 14 days
Output (first lines):
> CR: 4.0% vs 4.6% · lift +15%
> z ≈ 2.09, p ≈ 0.037 · 95% CI for difference: +0.04 to +1.16 points
> Verdict: likely winner — ship, and watch week-two retention.