/growth

Define and evaluate the experiment

Improve acquisition-to-revenue conversion by deciding what's worth testing, whether it can be answered at your scale, and what a result actually means once it exists. growth owns the feasibility gate, the design, and the interpretation — not the measurement machinery underneath it, and not the demand that fills the top of the funnel. The scope is funnel-wide: diagnosis, prioritization, feasibility, readout, activation, conversion optimization, retention, referral loops, pricing experiments, product-led growth, and quasi-experimental methods for when randomization isn't available. growth writes experiment designs, feasibility verdicts, readouts and requirements. It writes no production code, and it does not compose the variant it tests.

The default reader runs a self-serve product with real but limited traffic. Enough to test something, rarely enough to power a standard conversion-rate experiment the way a 200,000-visit playbook assumes. What remains at that scale is real: bolder tests, upstream metrics, and honest acceptance of a higher false-positive rate — never "you can't test."

The faceted router

SKILL.md is a router, not a script. Every request selects the smallest sufficient route: one primary job — the twelve below — combined with at most one base surface that reshapes how the job applies to how the product is actually bought, plus additive overlays when traffic is too small or a model runs the experiment. Read the selected references completely; load one or two, never the whole pack.

The twelve primary jobs

Each job is one reference, read fully only when its route is selected. Feasibility is the flagship because every other growth reference that proposes running a test assumes this file's gate has already been cleared. This is the whole funnel-wide surface, not a headline slice.

Full router table & invariants: SKILL.md.

Four surfaces + two additive overlays

One base surface, at most, reshapes every job for how the product is actually bought and used — the same conversion-test job resolves differently for a marketplace with spillover between arms than for a mobile app's IAP economics. The two overlays are additive — they stack on top of whichever base you picked, never replace it — and carry a distinct violet identity throughout this page, the same convention automation, operate, quality, data and marketing use for their own additive overlays.

Experiment design and feasibility — the flagship

Ask an agent whether a test is worth running and it either skips straight to statistics with nowhere to land, or straight to "run an A/B test" with no check on whether the traffic can support one. The gate comes first: can this specific comparison reach a power you'd trust, before a hypothesis or a flag exists. Skip it and you inherit the ecosystem's most common failure mode — even careful incumbent suites go straight from hypothesis to flag to metrics with no sample-size or MDE step in between. The teaching that matters most: halving the effect you want to detect quadruples the traffic you need — at a 5% baseline conversion rate, a +20% relative MDE needs 7,457 exposures per arm; +10% needs 29,826 (4.0×); +5% needs 119,303 (4.0× again). Metric choice alone can move that requirement two more orders of magnitude — a skewed metric like revenue-per-user can need ~114,000 where a low-skew metric needs ~1,550, on the same underlying change (Kohavi's skewness rule, KDD 2014).

CUPED will not rescue you here. Three real Microsoft experiments showed 45%, 52%, and 49% variance reduction with one week of pre-period data — but the same paper's own §5.2.3 reports revenue-per-user reduced by less than 5%, "due to the low correlation." It fails hardest exactly where growth spends most of its effort: new users have a covariate-metric correlation of 0.2–0.4 versus ~40% for existing ones (CUPED's own paper, Eppo, Statsig, and Netflix's KDD 2016 case study all agree), and retention sees small variance reduction for both new and existing users. The small-sample problem cannot be variance-reduced away precisely where activation and retention testing lives — sequential monitoring changes when you find out, not how much data the question needs.

What makes this different

Almost every individual fact in this pack exists somewhere else. What does not exist anywhere else is a pack that puts the growth surface and the validity layer in the same repo, and grades every benchmark it leans on rather than repeating it because it sounds measured — so the wedge is named as convergence and provenance discipline, not discovery.

The universal invariants

These govern every route, whichever references it loads.

• Say "open at scale," never "empty"Competent prior art exists (rampstack, GrowthBook, PostHog, coreyhaines31) and is cited respectfully; the gap is that the growth surface and the validity layer have never shared a repo.
• Cite `data`'s validity mechanics by name; never re-teach themSRM, CUPED's derivation, and peek-safe method selection live in `data`'s own reference. Growth's job is the other half of a seam `data` already drew in writing.
• Small-sample is a redirect, never a refusalTeach the three levers and the skip-the-test decision rule; label validity trade-offs honestly — switching to a proxy metric answers a different question, not a free win.
• Win rates never blendOnly Microsoft's ~1/3 (all four qualifiers, no published denominator) and Bing's 10–20% are first-party citable; every other famous win-rate figure is never-ship.
• Peeking is one curveState how many looks (K), the nominal α, and sidedness every time; a reconciled Armitage-to-Johari curve, never a bare percentage.
• CUPED always travels with its companion figureLead with Deng/Xu/Kohavi/Walker's 45/52/49%; carry the family's own honestly-labeled low-effect surface alongside it, never contradict it.
• The growth-vs-operate line is crispA standing threshold on the live system is `operate`'s; the same metric bound to a single tested change's decision is `growth`'s.
• Calculators are executable and unit-validatedNever a prose formula or a hand-built lookup table — two of three incumbent treatments checked this run were wrong, in opposite directions.
• A disclaimed figure is still a figureNever name a never-ship number in order to forbid it — state the mechanism and point at the primary source instead.

What a pass produces

Designs, verdicts and readouts are delivered as documents. Neither is delivered as production code. A growth artifact is incomplete unless it carries the metric and who picked it, the feasibility verdict and what it assumes about scale, every claim with a named proof source and its evidence tier, the OEC and guardrails if an experiment is being designed, the decision rule committed to before the result exists, what the result cannot answer, and the sibling packs it hands to or depends on.

The handoff seams

growth operates independently when invoked alone, and uses compatible upstream artifacts without silently overriding them. It is rarely terminal: handoff.md maps every seam between this pack and the rest of the family from growth's side — what it produces for each sibling, what it consumes, and the line it does not cross.

metric: <who picked it>
feasibility_verdict: <can this be
  answered, at what cost>
claims: [claim, named proof source,
  evidence tier]
oec_and_guardrails: <if an
  experiment is designed>
decision_rule: <committed before
  the result exists>
cannot_answer: <what the result
  does not prove>

The AARRR note. A practitioner who knows Acquisition → Activation → Retention → Referral → Revenue will notice this pack starts at Activation, not Acquisition — a deliberate seam, not a gap. AARRR spans three packs: marketing owns Acquisition — channels, positioning, demand; growth owns the experiments that improve conversion across the whole funnel — Activation, Referral loop design and math, and Retention experimentation; success owns Retention execution — the onboarding education and retention communications a growth test showed were worth sending. See growth-model-and-loops.md for the full reciprocation of marketing's own shipped note.

The rest of the family — each an independently installable pack with its own guide:

Start here

Install once. It's a plain SKILL.md router — no flags, no config, no fixed pipeline — so it activates on natural-language phrasing ("is a 3% lift worth testing at our traffic," "we peeked at this four times, is it safe to ship," "design a referral program and check the loop math") rather than a fixed command.

The same install runs on any Agent Skills host. Codex triggers with $growth; manual copy into any client's skills directory also works.