the-hypothesis-engine
Settle the minimum detectable effect before sizing anything. Sample size is a function of the effect you are willing to chase, so asking for it first prevents the common outcome: a twenty-week test powered to detect a lift too small to justify shipping. Ask what improvement would actually change a decision, and if the honest answer is a large one, the test gets much cheaper. Where the MDE implies a runtime longer than the decision can wait for, say the test is not viable and name the alternatives — a bigger change with a bigger expected effect, a proxy metric closer to the intervention, or a decision made without a test and reviewed later.
Vocabulary: Use "Experience" throughout, not "experiment" or "A/B test." This matches Intempt product terminology. When the reference file uses "experiment," translate to "experience" in all output.
Context
- Check for
.agents/product-context.md. If missing, ask the user to run/gtm:product-contextfirst. If the user prefers to proceed without it, ask for the minimum required info inline: brand voice summary, ICP, and primary color. - Read
references/bayesian-testing.mdfor statistical design patterns and Thompson sampling details. 2a. Read that file's Validity Threats section too. Those checks decide whether a result is readable at all, and they belong in the brief up front rather than being discovered after the experience has run: a confident number from a broken experience is worse than no number, because it gets shipped.