experiment
Installation
SKILL.md
Experiment Design — Orchestrator
Strategy — Step 4 of 4. Designs minimum viable tests with clear success/iterate/kill rules.
Core Question: "What's the cheapest way to learn if this works?"
Critical Gates — Read First
- Executable in ≤2 weeks. If the experiment takes longer than 2 weeks, it's not a minimum viable test. Simplify the design or switch test type.
- Success threshold MUST be a specific number. "Significant improvement" or "we'll know it when we see it" are not thresholds. "≥2.5% signup rate" is a threshold. Define before launching.
- Sample size MUST be checked. If insufficient for A/B, adjust: bigger swing, lower confidence, Before-After, or Pilot. Never rationalize "let's just go with what feels right."
- Guardrail metrics are mandatory. A test can "succeed" on the primary metric while degrading support tickets, error rates, or refund rates. Define at least 2 guardrail metrics with specific thresholds.