CSCustomer Success Skillsfor AI agents

30 agent skills · MIT licence

Customer Success Skills for AI agents

npx skills add gaintrace/customer-success-skills

Installs all 30 skills for Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Copilot and 20+ other clients.

Agent skills for churn risk, renewal forecasting, QBRs, expansion, escalations and board reporting — written to the bar a Chief Customer Officer would sign, and structured so an agent produces the same quality every time.

Claude Code — plugin
/plugin marketplace add gaintrace/customer-success-skills
/plugin install customer-success@customer-success-skills

Or say what you mean — “is northwind going to churn”, “prep me for the acme call”. Skills fire on the prompt, not the command name.

30Skills, each self-contained
7Signal families, fixed vocabulary
24Operating rules, cited by number
77Routing prompts in CI

The evidence standard

Most AI output in customer success fails the same three ways.

It invents a number to fill a gap, and someone repeats it to a customer. It silently omits what it could not check, so the reader assumes it was checked and clear. And it produces false precision — an “87% churn probability” from a rules-based score that has never been backtested against a single renewal.

MechanismWhat it does
Provenance tagsEvery number carries [system · field · date]. No tag, no number.
Three evidence tiersObserved / Inferred / Unknown, never blurred. Inferences state their rule and what would falsify them.
The Coverage LedgerEvery analysis ends with all seven signal families printed — including the ones with no data — plus a coverage percentage and the confidence cap it forces.
Confidence with entry criteriaHigh / Medium / Low / Insufficient, each with stated conditions. Confidence can never exceed what coverage permits.
Override floorsCommercial decisions — auto-renew off, notice served, data export — set a risk floor no healthy-looking aggregate can wash out.
No certainty language“100% accurate”, “guaranteed”, “will churn” are banned. Bands and confidence only.

“Never misses anything” here means something checkable: a fixed checklist walked every time, non-findings printed rather than dropped, blind spots declared, and no number that is not either computed or marked UNKNOWN — requires X.

Shared vocabulary

The seven signal families.

Every skill reasons over the same seven. Fixed vocabulary is what lets a risk assessment, a QBR, a forecast and a board report agree with each other.

Product usage & adoptionAre they still getting work done in the product?
Commercial & contractHave they taken an action that signals a decision?
Relationship & engagementDo we have a live, multithreaded relationship?
Support & reliabilityIs the experience costing them more than it returns?
Sentiment & VoCWhat have they actually told us?
Billing & paymentIs the money still flowing cleanly?
Firmographic & externalHas their world changed?

What it actually produces

Brief by default. Twenty lines, one recommendation, and the thing that would prove it wrong.

An excerpt from the worked example — a complete churn-risk run over a four-account renewal window. Read it for the mechanics, not the numbers.

churn-risk · briefInternal document
Renewal window · $1.12M assessed · $410k at risk (37%) · 2 accounts escalating regardless of score

1. Delta Systems — High Risk. $310k. Decide in 5 days.
A P1 has been open 19 days past SLA [Zendesk · #48211 · as-of 2026-08-26], seats were
cut 25% in March [Salesforce · Asset.Quantity], and invoices have slipped from 4 to 22
days late [Stripe · 180d]. Three separate pressures, one direction.
Do: Priya opens the exec call and names an engineering owner by 28 Aug.

2. Acme Corp — Critical. $148k. Decide in 6 days.
Auto-renew was switched off on 2 Aug [Salesforce · Contract.AutoRenew__c] and the
champion's email has hard-bounced since 11 Aug. The score is irrelevant — a decision
has been made.
Do: Marcus runs exec-to-exec outreach by 28 Aug, after verifying in 24h that
the flag change was not our own re-papering.

3. Beta Industries — Watch band, escalating anyway. $620k.
RevOps, who signed and hold the budget, went from 22 active users to 4 since May, while
the account total rose 18% because Marketing grew [Amplitude · distinct_users_30d].
Aggregate health reads green; the buying team has left.
Do: Jo meets Dana Osei (VP RevOps) by 10 Sept with RevOps' own numbers, not the
account total.

Confidence: high on Delta and Beta (7/7 families), high on Acme (6/7 — no VoC source).
What would change this: RevOps back above 12 weekly actives; Delta's P1 closing with
a committed date; confirmation that Acme's auto-renew change was ours.

Beta Industries scores 34.7/100 on the churn-risk scale — where higher is worse — with every aggregate metric healthy, and escalates anyway, because the team that holds the budget stopped using the product while the account total grew. The scoring is reproducible:python3 skills/churn-risk/scripts/risk_score.py …

The library

All 30 skills.

Each is a self-contained folder — SKILL.md plus its references, assets, scripts and evals. Take the whole library, or the two you want.

How they fit together

The common chains.

Skills hand off to each other on a shared schema, so the output of one is a legitimate input to the next. cs-context runs first and every other skill reads it.

Install

One command in Claude Code. A clone anywhere else.

Works with Claude Code, Claude, Codex, Cursor, Windsurf, Gemini CLI, Copilot and anything supporting the Agent Skills standard.

npx skills add gaintrace/customer-success-skills

The one that works everywhere. Add --skill churn-risk for a single skill, or -a claude-code to target one agent.

Claude Code — plugin
/plugin marketplace add gaintrace/customer-success-skills
/plugin install customer-success@customer-success-skills

Skills are namespaced — /customer-success:churn-risk — though you rarely type them.

Shell
claude plugin marketplace add gaintrace/customer-success-skills
claude plugin install customer-success@customer-success-skills
Clone
git clone https://github.com/gaintrace/customer-success-skills.git
cp -r customer-success-skills/skills/* ~/.claude/skills/

Or point your agent's skills directory at skills/.

First run

Start with cs-context. It writes.agents/cs-context.md, and every other skill reads it. Skip it and the skills will tell you the output is generic rather than guess your commercial model.

Set up customer success context for my company
Which of my renewals in the next 90 days are at risk, and why?
Prep me for the Acme call tomorrow
Where's my expansion revenue hiding?
What went wrong with the accounts we lost last quarter?

From skills to always-on

These skills do the thinking. They do not do the watching.

An agent session ends and the sweep goes stale the moment you close it. The signals that decide a renewal arrive between reviews — a champion's email starts hard-bouncing on a Tuesday, an auto-renew flag flips on a Friday, a P1 ages past its SLA over a weekend. Run the seven-family sweep by hand on Monday and it is accurate on Monday.

That gap is a job for a platform, not a prompt. Every skill closes with an Automate This note naming the manual cost it just incurred, so you can decide what is worth wiring up — with whatever tooling you already run.

GainTrace, a contributor to this library, builds one such platform — the column on the right is what it automates. The library is MIT and vendor-neutral: nothing here requires it.

The skills, manuallyWhat a platform automates
You reconcile CRM, billing, support, product analytics and email by hand, per accountPre-built connectors resolved into one live account timeline
You score risk once, in a session that endsContinuous scoring that ranks who needs attention today
You explain the score by hand so people trust itSignal-by-signal reasoning shown, never an opaque number
You find risk when you happen to lookRisk flagged ahead of the renewal call rather than at it
You write the play and then chase itPlaybooks that fire when risk signals cross a threshold

gaintrace.com ↗