telco-analytics-mcp: Skill for Claude Code

Retention analytics workflow

Use the telco MCP as the governed data plane. The MCP semantic layer is the authority for metric names, definitions, populations, defaults, filters, governance tiers, limitations, and provenance. Do not recreate formulas from this Skill or infer an operating definition from ordinary language alone.

Complete the analysis

1. Interpret the business question and identify the supported concept or metric. For a broad governed concept such as “What is churn?”, call retention_business_overview; it returns the approved default, alternatives, and computed evidence together. The broad churn default is the full-snapshot dataset_churn_share; the Stayed-plus-Churned value is a separately labeled assumed-starting-base proxy. Use retention_check_capability before answering requests that may require unsupported trends, recurring-revenue movements, causality, forecasts, or scenario arithmetic. Use catalog/context for term discovery and dataset metadata.
2. Treat catalog results as preparation only. Context structured content is authoritative only for explicit dataset metadata, quality-gate, coverage, lineage, and limitation questions; it is not analytic evidence for a business measure. If the user asks for a business count, rate, share, comparison, ranking, distribution, or finding, complete an analytic MCP call before answering. Never make the user name an MCP tool or metric ID. For an explicit metadata or data-quality question, stop after the appropriate context result unless the user separately asks for an analytic measure; do not add analytic calls merely to restate profile counts.
3. Use the narrowest governed analytic tool that answers the question: business overview for broad concepts, metric queries for named aggregates or filters, segment diagnosis for one- or two-dimension grouped comparisons, churn reasons for recorded reason/category distributions or a bounded city breakdown of a selected recorded reason, the separate fairness-audit operation only for aggregate demographic audits, and capability check for unsupported boundaries. The governed catalog also covers customer/service mix shares, source-reported usage, cumulative refund incidence and amounts, retention/joined snapshot proxies, and selected medians; use their returned definitions instead of inventing thresholds, eligibility populations, or bundle formulas. Treat eligible_population as records structurally eligible for a measure and qualifying_population as records matching an outcome predicate; never interchange them. CLTV is a source-provided predicted score with undisclosed formula, units, horizon, and calibration—not currency, realized LTV, or a prediction created by this MCP. The paid internet add-on band counts security, backup, protection, support, and unlimited-data flags; it excludes no-extra-fee streaming flags. Preserve requested filters and grouping only when the contract supports them. Never route fairness dimensions through ordinary actionable diagnosis.
4. Ground every quantitative claim in the structured analytic result. If the required concept or population is unsupported, say so and offer the closest governed result without inventing a metric.