Experiment Backlog — fewer bets, ranked, decided in advance

Most "testing" is activity wearing a lab coat. An experiment is only real once three things are written down before it runs: the hypothesis with its because, the smallest test that could falsify it, and the decision rule that fires when the result lands. This skill produces those — ranked, owned, and deliberately few.

Inputs

• The metric to move, plus anything already tried and how it went: $ARGUMENTS
• From marketing-brief.md if present: numbers that matter (north star and drivers), channels & funnel with stage conversion numbers, top pains, objections & beliefs, and past learnings. The weakest funnel stage is the best hypothesis generator.

Do this

1. Anchor on the target metric and find the weakest link feeding it in the brief's funnel numbers. Experiments aimed at the constraint beat experiments aimed at whatever is fashionable this month.
2. Generate 15–20 candidates, each written as: "We believe [change] will move [metric] because [observation or research]." No because, no backlog entry — a hunch gets upgraded by finding its evidence or gets cut.
3. ICE-score honestly and show the table:
   • Impact — how far the target metric moves if this wins.
   • Confidence — what evidence says it will. Enthusiasm scores a 1.
   • Ease — time and cost to a readable result. Sort by score. Note where two experiments would contaminate each other's read.
4. Design the top five as minimum-viable tests: the smallest version that produces a believable signal, what stays constant, the single success metric, and the rough sample or duration that signal needs — sanity maths, not false precision. Low-volume businesses read sales conversations and close rates, not dashboards.
5. Write the decision rule for each before it runs: "If X by [date or sample], scale it by doing Y. If not, kill it and start the next." No extending a test because the result hurt.
6. Set the operating cadence: how many run at once (one or two — parallel tests on the same audience muddy each other), the owner per test, and the weekly readout inside /marketing:weekly-review.

Output

The full scored backlog table, then the top five as test cards — hypothesis, minimum-viable design, success metric, sample or duration, decision rule, owner, start date. Log the five live cards in marketing-brief.md (create a Learnings & experiments section if none exists) so every future session knows what is in flight.

Rules

• Evidence over invention: confidence scores come from the brief or past results, never from how much the team likes the idea.
• One variable per test. Two changes that win together teach you nothing you can reuse.
• A test without a pre-written decision rule is a diary entry.
• Killing experiments is the system working. The backlog exists so stopping one thing never feels like stopping.
• Dead experiments get written down too. A kill nobody recorded is an experiment somebody reruns next quarter.