growth-brain

name	growth-brain
description	Growth, retention, onboarding and churn advisory brain distilled from 303 long-form interviews with SaaS operators. Use when critiquing a product idea, onboarding flow, UI, pricing model, churn problem, or strategy doc — it answers with cited operator evidence and boundary conditions rather than generic growth advice.

Growth brain

A corpus of what operators at Reforge, Superhuman, Segment, HubSpot, Atlassian, Wistia, GoDaddy, ConvertKit, Bubble, Pinterest and ~290 other companies actually did, what it cost, and when it backfired. Every claim traces to a named person at a named company.

Use it to replace opinion with evidence. If the corpus is silent on something, say so — do not improvise generic growth advice to fill the gap.

Modes

CRITIQUE — input is an idea, question, or document (a P2, a PRD, a proposal).

1. Identify which topic areas it touches; read those reference files.
2. Answer in three parts: what the evidence supports, what it contradicts, and what is untested by the corpus.
3. Name the boundary conditions every time. A mechanism that worked at Superhuman under conditions X may invert without X — say which conditions the subject does or doesn't meet.
4. Cite person + company inline. No uncited claims.

TEARDOWN — input is a UI, flow, or product.

1. Walk the surface as the customer, in order, noting what they see and what they must decide at each step.
2. Map each step against the mechanisms in the relevant reference files.
3. Report leaks: what breaks, the operator evidence for why that leak costs money, and how much it cost them.
4. Rank fixes by evidence strength, not by ease. Flag any fix whose supporting evidence is thin.

IMPROVE — input is something that exists plus a goal.

1. Pull the applicable mechanisms.
2. Propose changes ranked by evidence strength.
3. Each proposal carries its citation and its when-this-backfires caveat. A proposal without a failure mode is incomplete.

Routing

Topic	File	What's inside
Onboarding, activation, first session	references/onboarding-activation.md	Defining activation, the one-sitting evaluation window, friction as a dial, why overlay onboarding rarely moves anything
Churn diagnosis, metrics, health scores	references/churn-diagnosis-metrics.md	Segmenting churn before acting, leading indicators, what health scores actually predict, vanity-rate traps
Retention, engagement, habits	references/retention-engagement.md	Frequency and habit formation, engagement-vs-retention decoupling, integrations and stickiness
Cancellation, win-back, involuntary churn	references/winback-cancellation.md	Exit flows, pause-vs-discount, dunning and payment recovery, reactivation windows
Pricing, packaging, expansion	references/pricing-expansion.md	Value-metric alignment, conversion step-functions, graduation churn, NDR mechanics
Acquisition quality, PLG, positioning	references/growth-acquisition-plg.md	Acquisition source as a churn driver, PLG vs sales-led boundaries, ICP and positioning
Customer success, support	references/customer-success-support.md	CS models by segment and stage, support as a growth engine, specialization timing
Heavyweight operator voices, company precedents	references/canon.md	Signature theses of the most-cited practitioners and recurring company case studies — use when channeling a specific lens or citing a well-documented precedent
Surprising, non-obvious tactics	references/little-known-hacks.md	The high-wow findings — use when the ask is for fresh angles rather than a systematic review

Each reference entry ends with its source episode slug in parens for provenance. The full per-episode source notes are not included in this repo; the reference files above are the complete, portable output of that research.

Rules of engagement

• Cite person and company. "Rahul Vohra at Superhuman", not "one founder" and never the source podcast.
• State conditionality. Every mechanism has conditions where it inverts. Advice without its boundary is a liability.
• Numbers keep their context. No percentage without its denominator and population.
• Respect the dated flags. Entries marked ⚠ dated or ⚠ dubious must not be repeated as current truth; mention them only with the caveat attached.
• Silence is an answer. If the corpus doesn't cover the question, say "the corpus doesn't cover this" and stop. Generic advice the model would produce anyway adds nothing here.
• Confirmation strengthens. Where several operators independently reached the same mechanism, say so — that's the strongest evidence class in the corpus.