growth-experiments— agent skill

When to use

When a founder has traffic or early users and needs a disciplined way to find what moves the needle — instead of random tactics.

Inputs to gather

1. The one metric that matters most right now (activation, signup→paid, retention, referral).
2. Current funnel numbers (even rough) and the biggest drop-off step.
3. The primary acquisition channel and weekly bandwidth for experiments.

Process

1. Find the bottleneck stage in the funnel (acquisition → activation → revenue →

retention → referral). Run experiments where the leak is biggest, not where it's easy.

1. Generate 8-12 experiment ideas tied to that stage, each as a hypothesis:

"If we [change], then [metric] improves because [reason]."

1. ICE-score each: Impact, Confidence, Ease (1-10 each); score = average. Sort desc.
2. Pick the top 1-3 that fit this week's bandwidth. Define the success threshold and

minimum sample size before running — so the result is unambiguous.

1. Run the weekly loop: Mon pick & define → run → Fri read result → keep / kill /

double-down → log the learning. Compound the wins.

1. Channel focus: go deep on one channel until it plateaus before opening a second —

distribution beats spreading thin.

Output

• Funnel snapshot with the bottleneck stage flagged.
• Experiment backlog table: hypothesis · metric · ICE (I/C/E + score) · effort.
• This week's 1-3 experiments with success threshold and sample size each.
• Weekly loop template to reuse, with a running learnings log.

Quality bar & pitfalls

• Good = every experiment is a falsifiable hypothesis with a pre-set success bar.
• Score honestly — Confidence should reflect evidence, not enthusiasm.
• Avoid: vanity metrics (raw signups over activated users); running 10 things at once so

nothing is conclusive; chasing new channels before exhausting the working one; no learnings log (you'll re-run dead experiments).