Purpose
Turn a target metric and product context into a backlog of 12 rigorously structured experiment hypotheses — each with a causal mechanism, defined success and guardrail metrics, and a feasibility check — so the team tests high-leverage ideas in the right order instead of shipping random tweaks and hoping the number moves.
When to Use
Invoke this skill when the user asks to:
- Generate hypotheses or ideas for A/B tests to move a specific metric
- Build or refill an experiment backlog for the next quarter or sprint
- Decide what to test next when a KPI (activation, conversion, retention, ARPU) is stuck
- Convert a rough product idea into a properly formed, testable hypothesis
- Get expected mechanisms and success criteria for an experiment they already have in mind
- Prioritize a list of competing test ideas by expected impact and effort
Inputs Required
Before starting, collect:
- Target metric: The single primary metric to move (e.g., signup-to-activation rate, checkout conversion, D7 retention, trial-to-paid)
- Current baseline: The metric's current value and, if known, its trend
- Product context: What the product is, the surface being tested (onboarding, pricing page, checkout, feature X)
- Audience/segment: Who the experiment targets (new users, free tier, mobile, specific cohort)
- Traffic volume: Approximate weekly users or sessions on the tested surface (drives sample-size feasibility)
- Known constraints: Engineering effort ceiling, brand/legal guardrails, or ideas already tried and failed
- Guardrail metrics to protect: Metrics that must not degrade (e.g., revenue, support tickets, unsubscribe rate)
If baseline or traffic volume is missing, estimate a plausible range, state the assumption, and flag which hypotheses become infeasible at low traffic.
Instructions
Generate 12 Hypotheses Across 6 Intervention Levers
Produce 2 hypotheses per lever. Every hypothesis must follow this exact structure:
- Hypothesis statement: "If we [specific change], then [target metric] will [direction] because [causal mechanism]." — one sentence, no hedging.
- Mechanism: The behavioral or psychological reason the change should work (e.g., reduced cognitive load, loss aversion, faster time-to-value). 1–2 sentences.
- Primary metric: The exact metric this test moves, matching the target.
- Guardrail metrics: 1–2 metrics that must not regress.
- Expected effect size: A realistic relative lift range (e.g., +3–7% relative), grounded in the size of the change — not wishful.
- Feasibility flag: Given traffic volume, note whether the sample size is achievable in a reasonable window (2–4 weeks) or whether the effect is too small to detect at current traffic.
- Effort: Low / Medium / High engineering effort.
Lever 1 — Friction Reduction Remove or shorten steps between intent and outcome.
- Target: fewer form fields, fewer clicks, removed gates, autofill, progressive disclosure
- Strong for activation and conversion metrics where drop-off is step-based
- Mechanism basis: cognitive load, effort-to-reward ratio
Lever 2 — Motivation & Value Framing Change what the user believes they get, not what they do.
- Target: headline copy, benefit emphasis, social proof placement, outcome-first messaging
- Strong when drop-off happens before action, not during
- Mechanism basis: perceived value, relevance, aspiration
Lever 3 — Timing & Sequencing Change when a prompt, ask, or step appears.
- Target: moving a paywall, delaying an upsell, front-loading the aha-moment, re-ordering onboarding
- Strong for retention and monetization metrics
- Mechanism basis: readiness, momentum, time-to-value
Lever 4 — Incentive & Commitment Mechanics Add a reason to act now or a small commitment that compounds.
- Target: limited offers, progress bars, streaks, tiered rewards, trial extensions
- Strong for conversion and repeat-engagement metrics
- Mechanism basis: loss aversion, commitment consistency, goal gradient
Lever 5 — Personalization & Relevance Make the experience match the user's context or segment.
- Target: segment-specific defaults, recommended paths, tailored copy, use-case selection
- Strong when a broad audience has divergent needs
- Mechanism basis: relevance, self-selection, reduced choice overload
Lever 6 — Trust & Risk Reduction Lower the perceived cost of proceeding.
- Target: guarantees, security signals, transparent pricing, reversibility ("cancel anytime"), reviews
- Strong for high-consideration or payment-adjacent steps
- Mechanism basis: risk aversion, uncertainty reduction
Prioritize the Backlog (ICE + Feasibility)
Score each of the 12 hypotheses and rank them:
- Impact: expected effect size on the target metric (1–5)
- Confidence: strength of the mechanism and any prior evidence (1–5)
- Effort: inverse of engineering cost (1–5, where Low effort = 5)
Compute an ICE score, then apply a feasibility override: any hypothesis flagged as undetectable at current traffic drops in rank regardless of ICE, with a note on the traffic needed to make it viable.
Present the top 5 as a recommended test sequence, noting which tests can run concurrently (non-overlapping surfaces) and which must run sequentially (shared surface or interacting effects).
Output Format
- Structure: 6 lever sections (2 hypotheses each, fully structured) → ICE-ranked table of all 12 → top-5 recommended test sequence
- Length: 1,000–1,400 words
- Format: Structured bullets per hypothesis; a ranking table; a numbered sequence for the top 5
- Sections: Assumptions/inputs recap → 6 levers with 12 hypotheses → prioritization table → recommended test sequence
- Tone: Senior PM — precise, mechanism-driven, skeptical of vague wins; no hype, no "should improve engagement" hand-waving
Quality Checklist
Before delivering output, verify:
- [ ] Exactly 12 hypotheses generated — 2 per lever (count them)
- [ ] Every hypothesis uses the "If we X, then metric Y will Z because [mechanism]" form with a stated causal mechanism
- [ ] Every hypothesis names a primary metric that matches the target and at least one guardrail metric
- [ ] Expected effect sizes are ranges grounded in the change, not single optimistic numbers
- [ ] Feasibility flags reflect the traffic volume provided — small-effect tests at low traffic are marked undetectable
- [ ] Top 5 sequence notes which tests can run concurrently vs. must run sequentially
Notes
- The most common failure is a hypothesis with no mechanism — "change the button color" is not a hypothesis, "increase contrast to make the primary action more salient and reduce decision hesitation" is.
- Guardrail metrics matter most for incentive and pricing tests, where the target metric can improve while revenue or trust quietly degrades.
- At low traffic, prefer high-impact, structural changes (friction removal, timing) over micro-copy tweaks — small effects are unmeasurable and waste test slots.
- Encourage running one hypothesis per surface at a time unless the tests are genuinely independent; overlapping tests contaminate attribution.
How to use this skill
This file is a Claude Code skill in Anthropic's standard format — markdown with YAML frontmatter at the top.
Installing in Claude Code:
~/.claude/skills/<folder-name>/SKILL.md— auto-loads in every session anywhere on your machine (recommended for skills you'll use across multiple projects)<your-project>/.claude/skills/<folder-name>/SKILL.md— auto-loads only when working in that project (recommended for project-specific skills)
Restart Claude Code or open a new session to register.
Using in Claude.ai, ChatGPT, Gemini, or other AI tools:
- Drag this file into a new chat, or copy and paste its contents
- Send your task or request as your next message — the AI will apply the skill where it fits
For Claude Code skill details: https://docs.anthropic.com/en/docs/claude-code/skills