A Claude Skill that takes a churned account and produces a structured root-cause analysis: triggering event, contributing factors, deviation from the success plan, missed signals, and a prevention recommendation. Output is consistent across CSMs so RevOps can aggregate root causes quarterly without recoding free-text notes.

What you’ll need

Claude.ai or Claude Code
HubSpot or your CRM with the churn record (close-lost reason, churn date)
Gainsight or your CS platform for health scores and timeline events
Optional: Gong call snippets from the last 180 days
A churn taxonomy (5-10 root cause categories your team has agreed on)

Setup

Define the taxonomy. Edit references/churn-taxonomy.md with your top-level categories (e.g., product-gap, champion-departure, pricing, consolidation, service-failure). The Skill forces every analysis into one primary and up to two contributing categories from this list.
Install the Skill. Drop churn-analysis.skill into ~/.claude/skills/. Set HUBSPOT_TOKEN and GAINSIGHT_TOKEN in your environment.
Run on the churn. analyze_churn(account_id="...", churn_date="2026-04-15"). The Skill pulls the timeline, calls, and notes, then produces a one-page analysis.
Review with the CSM. The CSM owning the account corrects any factual errors and confirms the root-cause categorization. RevOps stores the final document in a shared Notion or Drive folder.

How it works

The Skill builds a 180-day timeline from CRM events, health-score changes, support cases, and Gong calls. It looks for inflection points — a sudden health drop, a sponsor change, a missed QBR — and ranks them by proximity to the churn date.

Then it runs a two-pass Claude analysis. The first pass extracts evidence: quotes from Gong calls, support ticket patterns, usage drops. The second pass classifies the churn against your taxonomy and produces the structured output. Crucially, the Skill refuses to assign a root cause if evidence is thin — it returns “insufficient data” rather than guessing, which keeps the aggregated quarterly view honest.

The prevention recommendation is the actionable part: what specific process change, alert, or playbook update would have caught this earlier.

Watch-outs

Hindsight bias. It’s easy to construct a clean narrative after the fact. Cross-check the Skill’s output against the CSM’s lived experience — if the CSM disagrees, their judgment wins.
Taxonomy creep. Don’t let the taxonomy grow past 10 categories. The whole point is aggregation; if every churn is unique, you can’t see patterns.
Champion-departure attribution. Easy to over-attribute to this category. Require evidence (a LinkedIn departure date or a CRM contact-change record) before the Skill assigns it.

Stack

HubSpot — churn record and contact history
Gainsight — health scores, timeline events, success plans
Claude — timeline synthesis and root-cause classification