菜单
Classify free-text cancel reasons into canonical buckets (price, feature_gap, temporary, quality, business_end, duplicate_account, unknown) using the YAML taxonomy. Run after cohort is built, before the strategy engine.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Show Annual Recurring Revenue with QoQ and YoY trend and a growth trajectory. Use when the user asks about ARR, annual recurring revenue, annual run rate, or year-over-year revenue growth. Do NOT use for MRR, NRR, or monthly breakdowns.
Identify overdue invoices most likely heading toward write-off, while a save is still possible. Use when the user asks about bad debt, write-off risk, invoices unlikely to be recovered, or wants to see which AR is at risk of becoming bad debt. Do NOT use for general overdue lists or collections prioritisation.
List subscriptions whose payment card expires in the next 30 or 60 days, ranked by MRR at risk. Use when the user asks about expiring cards, cards about to expire, or wants to prevent involuntary churn from payment method expiry. Do NOT use for current payment failures or dunning analysis.
Show the full product catalog: all products, their plans with pricing and billing frequency, addons, and one-time items. Use when the user asks what plans are available, wants to see the product catalog, pricing, or available addons and items. Do NOT use for subscription status, customer-specific data, or coupon lookup.
Decompose churn into voluntary (customers who chose to leave) vs involuntary (failed payments) and identify which plans and products are driving losses. Use when the user asks why customers are churning, churn composition, voluntary vs involuntary split, or which products have the highest cancellation rate. Do NOT use for MRR, NRR, or retention rate questions.
Daily AR worklist: top 20 overdue invoices ranked by recovery probability (outstanding amount × payment-behaviour score). Use when the user asks who to chase for payments today, wants a collections prioritisation, or needs a daily AR action list. Do NOT use for aggregate overdue totals or bad-debt analysis.
| name | cancel-reason-classifier |
| description | Classify free-text cancel reasons into canonical buckets (price, feature_gap, temporary, quality, business_end, duplicate_account, unknown) using the YAML taxonomy. Run after cohort is built, before the strategy engine. |
Turns cancel_reason_raw into a canonical cancel_reason_bucket via keyword matching against config/retention-matrix.yaml (reason_taxonomy key). Deterministic, zero LLM tokens, re-runnable.
| Field | Description |
|---|---|
cancel_reason_bucket | Canonical bucket from YAML taxonomy |
cancel_reason_match | Keyword that triggered the match; null for unknown |
Writes cohort-<type>-classified.json next to the input file.
1. Run the classifier script:
python <skill-path>/scripts/classify_reason.py \
--cohort <workspace>/retention-runs/<DATE>/cohort-<type>.json \
--matrix <workspace>/config/retention-matrix.yaml \
--output <workspace>/retention-runs/<DATE>/cohort-<type>-classified.json
Script logic: lowercases cancel_reason_raw, walks YAML buckets top-to-bottom, first keyword match wins. Prints per-bucket counts + up to 10 unknown examples to stdout.
2. Handle unknowns:
unknown (strategy engine has fallback rules).10% unknown → prefer Option A: add keywords to the YAML taxonomy and re-run (every future run benefits). Use Option B (inline LLM classification) only if reasons are too varied for keywords — set
cancel_reason_match: "llm_fallback"for audit.
3. Report per-bucket counts to user and confirm output path.
cancel_reason_raw → treat as unknown.