# growth

> Improve acquisition, activation, retention, referral, and revenue through honest
> experimentation. Installs as an agent skill.

Ask an agent whether a test is worth running and it either skips the statistics (activation,
retention, loops and PLG advice with no validity layer underneath) or skips the growth surface
(rigorous statistics with nowhere to land). `growth` is both in one pack: the feasibility gate,
experiment design and interpretation, funnel and cohort diagnosis, activation, conversion
optimization, retention, referral loops, monetization and pricing tests, product-led growth, and
quasi-experimental methods for when randomization isn't available. It is a **faceted router, not a
pipeline**: each request selects exactly one of 12 primary jobs, at most one of 4 base surfaces
(self-serve SaaS is the default, and an assumed surface is stated out loud), plus additive overlays
for small-sample traffic and agent-run experiments — so two references get read fully out of
nineteen, never the whole pack. The flagship job, `experiment-design-and-feasibility`, answers what
the rest of the field avoids: can this question be answered at your traffic, and what would a
"significant" result actually be worth — the Ambition Tax, where a small sample forces a bigger bet
and a bigger bet carries a lower prior, and the two effects multiply rather than cancel. The default
reader runs a self-serve product with real but limited traffic — rarely enough to power a standard
conversion-rate test the way a 200,000-visit playbook assumes. `growth` writes experiment designs,
feasibility verdicts, readouts and requirements — no production code. Claude Code is the reference
host; it's plain Agent Skills format (agentskills.io), so it installs on Codex, Cursor, Antigravity,
opencode, Grok Build, and Hermes too.

## Install

- npm/skills.sh: `npx skills add gabros20/growth-skill -g -y`
- Clone: `git clone https://github.com/gabros20/growth-skill`
- Manual copy: `cp -R skills/growth ~/.claude/skills/growth`

## Docs

- [README.md](https://raw.githubusercontent.com/gabros20/growth-skill/main/README.md): overview, install, examples, and repository map.
- [skills/growth/SKILL.md](https://raw.githubusercontent.com/gabros20/growth-skill/main/skills/growth/SKILL.md): runtime router, workflow, artifact contract, and completion rules.
- [SOURCES.md](https://raw.githubusercontent.com/gabros20/growth-skill/main/SOURCES.md): source attribution, licensing rule, and the numbers this pack refuses to ship.

## Reference inventory (`skills/growth/references/`)

Primary jobs — exactly one per request:

- `experiment-design-and-feasibility.md` — **flagship**. The feasibility gate; the derived power table against vendor floors; the metric-skew rule; the Bayes posterior on a "significant" winner; the Ambition Tax; where CUPED fails.
- `growth-model-and-loops.md` — the shared vocabulary. Loops vs funnels, the Racecar framework correctly attributed, Verna's Five Laws, the K-factor/cycle-time math the ecosystem never built, the AARRR split reciprocated with `marketing`.
- `funnel-and-cohort-diagnosis.md` — where a funnel is actually leaking. The 3-way retention definition — N-day, rolling, survival.
- `opportunity-and-prioritization.md` — what to work on next. ICE read skeptically as throughput tooling, not evidence.
- `experiment-readout-and-learning.md` — reading a finished experiment. The one-curve peeking reconciliation, the winner's-curse haircut at readout, Twyman's law as a Bayesian prior.
- `activation-and-onboarding.md` — activation and onboarding. Why CUPED fails for new users; "aha moment" labeled folklore with no traceable origin.
- `conversion-optimization.md` — a conversion-rate test on a page, form or button. The red-button-on-blue-page external-validity lesson; the copy→placement→color test-order rule.
- `retention-and-resurrection.md` — churn reduction, resurrection and win-back. The Duolingo streak specimen, fully worked and resolvable.
- `referral-and-product-loops.md` — designing a referral program. The loop math the ecosystem lacks, applied to a live design.
- `monetization-and-pricing-experiments.md` — pricing and packaging tests. Booking's own pricing-test refusal; Van Westendorp's provenance; the subscription dark-patterns table.
- `product-led-growth.md` — evaluating a PLG motion. Five-field benchmark-provenance discipline; the Sean Ellis 40% test with its own generalizability caveat.
- `quasi-experiments.md` — a causal question when randomization isn't available. Precondition checklists in place of numeric floors.

Base surfaces — at most one, reshaping how a job applies:

- `surface-selfserve.md` — **default**. Anyone can sign up or trial without talking to a human.
- `surface-b2b-sales-assisted.md` — a rep, demo or procurement gates the deal.
- `surface-mobile-subscription.md` — IAP economics; the RevenueCat-layer dataset.
- `surface-marketplace-network.md` — interference/SUTVA risk; diagnose the mechanism first.

Additive overlays — stack on a base surface, never instead of one:

- `overlay-small-sample.md` — traffic or users too small for standard power. A redirect, never a refusal.
- `overlay-agentic.md` — a model designs, runs or reads an experiment.

Seams:

- `handoff.md` — every boundary with the rest of the family, from growth's side, plus the AARRR note (acquisition in `marketing`, experiments here, retention execution in `success`).

## Assets (`skills/growth/assets/`)

- `power_calc.py`, `srm_check.py`, `peeking_table.py`, `skew_check.py` — runnable, self-testing calculators, each validated against a published anchor at import time (PostHog, Booking.com, Fabijan et al., the Armitage table, Bing's post-erratum skewness table).

## Optional

- [Visual guide](https://growthskill.vercel.app): the router, jobs, surfaces, flagship, invariants, and seams on one page.