ApexCloud Retention Ops
Use this skill for any task that turns ApexCloud operational data into a strict JSON artifact.
First pass
- Read the prompt first.
- Read the task's answer template from its input payloads next. Treat it as the schema oracle.
- Identify the report family before you fetch data.
- Use the task environment URL provided with the exercise and the allowed ApexCloud endpoints only.
Source map
- Read
references/source-map.mdfor the endpoint map and field rules. - Use it when you need to remember which source provides ARR, A/R aging, ticket quality, monthly revenue, or churn data.
Core workflow
- Copy the exact top-level structure from the answer template.
- Pull only the data needed for the requested report.
- Reconcile names, dates, account IDs, and periods against the prompt before you compute anything.
- Apply the report-specific rules below.
- Verify formatting, ordering, and enum values against the template.
- Return JSON only. No commentary, markdown, or code fences.
Report playbooks
Renewal-risk queue and retention board
- Use only the selected account IDs.
- Pull account profile, monthly metrics, tickets, NPS, billing snapshot, and A/R aging data.
- Use the billing snapshot at the requested as-of date for current ARR or exposure. If a snapshot is missing, fall back to the account profile only as a secondary check.
- Treat overdue balance as the sum of the 61_90 and 90_plus A/R buckets at the requested as-of date.
- Count clean tickets by excluding duplicate and spam tickets.
- Use the latest non-retracted NPS in the requested window.
- Rank by the strongest retention risk signals first: overdue receivables, renewal proximity, NPS decline, SLA degradation, usage decline, then tenure or lifecycle context.
- Use the prompt's controlled action and reason labels exactly. Do not invent labels.
- If the prompt asks for a "standard retention board order," preserve the risk order implied by the prompt and break ties deterministically with exposure or urgency.
QBR packet
- Build the monthly series from the requested quarter or month range.
- Use
recognized_revenue, clean ticket counts, SLA compliance, and NPS from the account-level metrics, tickets, and NPS data. - For ticket counts, exclude duplicate and spam tickets.
- For SLA compliance, compute the share of clean tickets whose first-response SLA was met.
- Use the earliest month as the default tie-breaker for tied highlight fields unless the prompt says otherwise.
- Treat ticket trend as improving when clean ticket volume decreases over the requested period, worsening when it increases, and flat when it is unchanged.
- Choose agenda topics only from the vocabulary in the prompt.
Receivables and pipeline review
- Pull overdue balances from
/api/finance/ar-agingat the requested as-of date. - Link A/R customers to CRM accounts by account_id first, then by normalized legal name, display name, or aliases.
- Mark the row
linkedonly when the CRM match is real and the account_id can be resolved. - Sort overdue follow-ups alphabetically by
customer_nameunless the prompt says otherwise. - Sum open opportunities that fall inside the requested window, and separate won/lost/open counts exactly as the prompt asks.
- Pull HR and event context from the quarter-level summary endpoints requested in the prompt.
Churn validation and candidate ranking
- Treat
train.csvandvalidation.csvas labeled data andcandidates.csvas unlabeled scoring input. - Validate the dataset from the files themselves before ranking candidates.
- Keep preprocessing identical across train, validation, and candidate files.
- Use a deterministic classifier or scorecard; do not hand-wave probabilities.
- Report probabilities with the precision requested by the prompt, and rank candidates strictly by probability descending.
- Keep the top-N list to the exact candidate set named in the prompt.
Formatting rules
- Match the template's key order, nesting, and nullability.
- Use controlled enum values exactly as shown in the prompt or template.
- Use integers for counts and risk scores, 2 decimals for currency, 1 decimal for percentages, and 3 decimals for churn probabilities when requested.
- Do not round early. Round once at the end.
- Keep lists sorted exactly as the prompt requests.
Final check
- Confirm every required top-level key is present.
- Confirm the output is valid JSON with no trailing text.
- Confirm no task-specific example values leaked into the reusable instructions.