/success

Hold the relationship, honestly

Keep the customer relationship after the sale, and be honest about what you can actually know about it. success owns the post-purchase jobs — the onboarding program, help content, the support process, AI-support policy, health and risk signals, voice-of-customer interpretation, retention execution, renewals and expansion, and the customer's side of an incident — and it owns the epistemics that come with them: which of those signals can carry a decision, and which cannot. success writes programs, policies, taxonomies, and workflows. It writes no production code, no experiments, and no numbers it cannot source.

The default reader is a builder with users, not a CS team with a book of accounts. So the front door is a gate, not a playbook — because no source in the CS canon ever says when a distinct success function becomes worth having. Where the enterprise canon still teaches something useful, this pack lifts its epistemics — input-expiry rules, segmented thresholds, treating missing data as risk, tie-breakers on a category list — and declines its machinery — CSM ratios, QBR cadence — with reasons stated, never by omission.

The existence gate — the front door

Before a single onboarding plan, health score, or escalation policy gets written, this pack asks whether a distinct success function is worth building yet. No source in the CS literature — Gainsight's own book, the CSM-ratio literature, any vendor playbook — states when that becomes true; the whole canon assumes the function already exists and only asks how to staff it. The Existence Gate uses the one practitioner sentence that makes it checkable, from Arvid Kahl, a solo founder: "Customer Success (compared to mere customer support) is so much more important when your revenue scales with your customer's revenue."

The ten primary jobs

Each job is one reference, read fully only when its route is selected. Health and risk signals is the flagship because it is the file every red score, every renewal risk, and every churn-save motion eventually has to answer to. This is the whole post-purchase surface, not a headline slice.

Full router table & invariants: SKILL.md.

Three surfaces + one additive overlay

At most one base surface reshapes every job for how the product is actually run — the same onboarding job resolves differently for a founder answering their own email than for a CSM-staffed account. The agentic overlay is additive — it stacks on top of whichever base you picked, never replaces it — and carries the violet identity throughout this page, the same convention automation, operate, quality, data, marketing, and growth use for their own additive overlays.

Health and risk signals — the flagship

Ask an agent to build a churn model and it either ships a color with no math behind it, or a number with no source. This pack checks first: can a score or model actually predict anything, and is the churn event even observable in the first place — before either ships. Even the most transparent public customer-success program in the research, GitLab's own handbook, does not ask its health score to predict churn: the score is one artifact, prediction is two separate internal models, and neither publishes an accuracy figure.

What transfers from the enterprise canon, at no cost: input expiry (show nothing rather than stale data), segmented thresholds (a measure switched off for a tier, not scored badly), treating missing data as risk rather than neutral, and a tie-breaker rule for every category boundary — none of it requires the CSM ratios or QBR cadence it usually ships wrapped in.

What makes this different

Customer-success advice is not scarce — it's fragmented. Real tools cover real pieces of this surface; none of them integrate, and almost none of them cite a source for the numbers they repeat. What doesn't exist anywhere else is a pack that puts the whole post-purchase relationship and the honesty layer underneath it in the same repo, and grades every figure it leans on instead of repeating it because it sounds measured.

The universal invariants

These govern every route, whichever references it loads.

• The gate runs before the apparatusA health score, playbook, or renewal workflow requested without justifying that success applies yet gets the gate first, and a named sibling for the meantime — never a silently invented artifact.
• Say "not published," never "not measured"No source discloses a health-score or churn-model accuracy figure against real outcomes. Absence of disclosure isn't proof of inaccuracy — but a reader can't verify what was never published.
• Ask the Observability Question before scoring anythingIn a contractual setting churn is observed on a known day; in a non-contractual one it isn't observed at all, and a contractual-shaped template is the wrong instrument.
• Score and prediction are different artifactsA health diagnostic and a churn prediction are not the same thing even at the org that invested most visibly in both. A score alone next to a renewal date is folklore; a score with a stated gap and a mandatory action is a ranking device.
• A category list without tie-breakers isn't a taxonomyTicket categories, churn reasons, and feedback reasons alike need a rule for the boundary case and a named default direction.
• An SLA is a promise to a specific personNever a copied template number. No response-time figure ships without "set this from your own observed capacity" framing.
• Grade every figure inlineMeasured-with-method > practitioner-report > vendor-claim, and carry the method with the number — or say "method unknown."
• 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.
• Always write "customer escalation"The bare word is loaded across four sibling packs and three senses inside support process itself.
• Lift the epistemics, decline the machinery — with reasonsA stakeholder asking for a CSM ratio or a QBR cadence gets a reasoned decline and the question it was trying to answer, never silence and never a borrowed benchmark.

What a pass produces

Gate verdicts, programs, taxonomies, policies, score audits, and workflows are delivered as documents. None is delivered as production code. A success artifact is incomplete unless it carries the gate verdict (or a statement that success already applies), every figure with its source and evidence grade, the seam naming which sibling owns the adjacent half, and what the artifact cannot support.

The handoff seams

success operates independently when invoked alone, and uses compatible upstream artifacts without silently overriding them. handoff.md maps every seam between this pack and the rest of the family from success's side — including three rows no sibling has drawn yet, because success ships after product, operate, and automation did.

gate_verdict: <applies /
  lighter-weight / not-yet>
job_and_surface: <which one,
  and any surface assumed>
claims: [claim, named proof source,
  evidence tier]
seam: <sibling pack + filename
  owning the adjacent half>
cannot_support: <what the
  artifact does not prove>

Three rows this pack originates. product, operate, and automation each already cede work to success in their own prose — but none carries a reciprocal success row, because success shipped after they did. This file states those seams from success's side: product hands success TTFV and pricing terms and receives support/satisfaction/churn evidence back; operate hands success the fact, impact, and timing of an incident, and success turns it into customer-facing comms, co-owned with product, never alone; automation's refund-approval flagship is a support workflow end to end — success owns the process and the refund policy, automation owns the approval gate.

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 ("do we need customer success yet," "audit our health score before we trust it," "draft our AI-support escalation policy") rather than a fixed command.

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