القائمة
ApexCloud Retention Operations — reusable rules and verified conventions for producing renewal-risk, QBR, receivables/pipeline, churn-model, and retention-board answers from the data API.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Reconcile MedBridge Sales Ops CRM tasks (B2B medical wholesale + implementation-services) against the remote Sales Ops API and emit the account-ready JSON answer. Covers QUOTE decision packages (EXW pricing, freight options, route risk, recommended mode, payment terms) and milestone RECONCILIATION packages (invoice/payment state, outstanding balance, revenue-recognition coverage, CRM follow-up routing, event/voucher linkage). Use whenever a task references MedBridge, quotes like Q-TR-*, RFQ-TR-*, opportunities like OPP-TR-*, milestones MS1/MS2/MS3, freight-quotes, revenue-journals, or the shared Sales Ops API base URL.
SCN_003 CRM support-operations solver. Resolve ticket batches (fixed-line diagnostics + gates), mobile contact-center triage, and enterprise export-complaint response packages against the shared support-console API. Apply exact metric thresholds, gate order, root-cause->team map, mobile action decision tree, and enterprise package derivation rules.
SCN_003 CRM support-console analyst. Resolve same-day ticket batches, mobile triage queues, mobile-data recovery worklists, and enterprise export-complaint response packages by querying the shared support-console API and applying fixed diagnostic thresholds, gating order, root-cause routing, and per-family decision trees.
Resolve CRM support-console tasks (SCN_003) — ticket-batch classification, mobile-data triage, and enterprise export-complaint response packages — against a remote support-console API. Use when a task asks to classify service tickets / queue snapshots, triage mobile data cases, or build an enterprise export-failure response package and references a support console API.
Resolve SCN_003 CRM support-console tasks: ticket-batch resolution, mobile contact-center triage, and enterprise export-complaint response packages. Drive the remote console API, apply diagnostic floors + gating order, and emit the per-family JSON answer.
Solve ApexCloud Retention Operations analytics tasks against the remote Retention Operations API. Encode and reproduce the exact business rules, data-hygiene conventions, policy codes, ranking/precision rules, churn-model spec, and per-archetype output schemas. Use for any CRM retention task (renewal risk queue / save plan, QBR metrics packet, receivables + pipeline operations digest, churn model validation + outreach ranking, high-touch retention action board / watchlist).
| name | reflect-3-attempt-02 |
| description | ApexCloud Retention Operations — reusable rules and verified conventions for producing renewal-risk, QBR, receivables/pipeline, churn-model, and retention-board answers from the data API. |
Reusable rules for answering the five task archetypes served by the ApexCloud Retention Operations API. Every rule below was checked against the train judge; conventions marked VERIFIED produced correct deterministic fields (train_003 reached 1.0; train_002 deterministic block correct). Items marked BEST-GUESS could not be fully recovered from scalar feedback and should be applied with the stated logic.
ENV_URL.txt (data API). Judge path from JUDGE_PATH.txt. All endpoints are read-only JSON./api/health, /api/accounts, /api/accounts/<id>, /api/accounts/<id>/metrics?start=YYYY-MM&end=YYYY-MM, /api/accounts/<id>/tickets, /api/accounts/<id>/nps, /api/billing/snapshots, /api/finance/ar-aging?as_of=YYYY-MM-DD, /api/opportunities, /api/hr/summary?quarter=YYYY-QN®ion=..., /api/events/performance?event=...&quarter=YYYY-QN, /exports/churn/{train,validation,candidates}.csv, /exports/account_metric_extract.csv.accounts endpoint returns {"accounts":[...]}; billing/snapshots returns {"snapshots":[...]} (key billing_arr, field posted: true, quarterly as_of = 03-31/06-30/09-30/12-31); ar-aging returns {"ar_aging":[...]} (buckets 1_30, 31_60, 61_90, 90_plus, keyed by customer_name = legal name, NO account_id field); opportunities have account_id, stage, close_date, amount, state.YYYY-MM-DD / YYYY-MM.current_arr = billing_arr from the posted billing snapshot whose as_of == the assessment date. Snapshots are quarterly; the assessment date (e.g. 2026-06-30) always matches a quarter-end as_of.account.crm_arr or account.billing_arr_current (the account's billing_arr_current equals the LATEST/2026-12-31 snapshot, not the assessment-date one). Example: globex_north assessment 2026-06-30 → snapshot billing_arr=1,176,600.70 (NOT billing_arr_current=1,188,000.00).model_checks.uses_billing_arr_source = true; tenure_risk_direction = "negative" (churn-model tenure coefficient is negative)./api/accounts/<id>/tickets rows filtered to the analysis period (created_date within [start, end]), EXCLUDING is_spam=true, is_duplicate=true, and status="cancelled". Ticket statuses observed: open, closed, cancelled.clean_ticket_count field (risk archetypes) AND the support_tickets field (QBR archetype): use the cleaned /tickets count, NOT metrics.support_ticket_count (which is the raw count including spam/dup/cancelled). Example: globex 2026-06 raw=3, clean=1. (/exports/account_metric_extract.csv exposes a clean_ticket_count column that equals the cleaned endpoint count — use it to cross-check.)sla_compliance (1 dp) comes from the account metrics endpoint per month; source label sla_report./api/accounts/<id>/nps, drop retracted=true and score == -1 (invalid), then take the latest by response_date within [start, end].nps_score in the metrics endpoint already equals the latest valid /nps response for that month — either source works. Source label nps_survey.overdue_balance = ar_aging.61_90 + ar_aging.90_plus for the as_of date given in the prompt (the "older aging buckets"). 1_30 and 31_60 are current, NOT overdue.customer_name (legal name); match to accounts by legal_name.customer_name to a CRM account by exact legal_name equality only. link_status="linked", account_id=<id>.link_status="unlinked"). Verified unlinked examples: "Globex North Subsidiary LLC" (vs "Globex North Holdings LLC"), "Valence Payment Services Canada" (vs "...LLC"), "North Star Finance Services" (vs "Northstar Finance Group Inc."), "Quartz Insurance Claims Ltd.", "Riverbend Bank Foundation". Aliases in account_aliases are NOT used for linking.overdue_followups sorted by customer_name ascending.close_date falls in the analysis date range. "Region: all regions" → do NOT pass region=all (returns empty); use the unfiltered /api/opportunities list and filter by close_date client-side.Closed Won → won; Closed Lost → lost; all other stages (Discovery, Prospecting, Negotiation, Proposal / state="open") → open.won_revenue = sum amount of won; open_pipeline = sum amount of open; win_rate_pct = won/(won+lost)*100 (1 dp). top_open_product_line = product line with the largest total open amount.region=all (and region=na/emea/apac/latam) return EMPTY; region=North America is also rejected by the filter. Query /api/hr/summary?quarter=YYYY-QN with NO region to get all 4 regional rows, then sum across the 4 regions (North America, EMEA, APAC, LATAM). hr_headcount = sum headcount; unpaid_claims_total = sum unpaid_claims_amount./api/events/performance?event=<id>&quarter=YYYY-QN returns one row; use event_orders and event_revenue directly (NOT completed_orders/product_revenue).expansion_offset or clean_billings, which are reducing/informational).overdue_receivable — overdue_balance > 0.low_tenure_high_churn — contract_tenure_months < 24.sla_degradation — mean monthly sla_compliance < 92.nps_drop — latest valid NPS < 60.usage_decline — last-month product_usage < first-month × 0.97 (>3% drop).renewal_window — renewal_date within [assessment−30d, assessment+90d] OR lifecycle_status == "renewal_risk".expansion_offset — open expansion opportunities exist in the period (sum > 0); informational, reduces effective risk.clean_billings — no overdue and healthy billings (informational).expansion_offset applies and the account is borderline.risk_score and the precise set/level per account could NOT be recovered from scalar judge feedback (train_001/train_005 stayed low). Treat the ranking+levels as the least-certain part; the deterministic per-account fields (current_arr, overdue_balance, latest_nps, clean_ticket_count, expansion_pipeline) ARE reliably computable.collections_followup, technical_recovery, renewal_save, executive_qbr, nurture_monitor, no_action.collections_followup; critical+Strategic (no overdue) → executive_qbr; renewal_window/renewal_risk → renewal_save; sla_degradation/usage_decline → technical_recovery; low risk → nurture_monitor; clean → no_action.next_touch_due_date = the date for the account's chosen primary_action; followup_calendar echoes the full mapping. For receivables-only tasks the follow-up due date is the single date stated in the prompt (e.g. 2026-10-15).primary_action/reason_codes did NOT move the judge score on the block-scored board archetype (train_005); the board appears scored as a block requiring correct order+levels+actions+reasons simultaneously. Prioritize getting the ORDER and risk_levels right.review_owner, needs_technical_signoff, agenda_topics) are best-guess: review_owner likely customer_success for healthy accounts / solutions_engineering when a technical SLA dip (<95) warrants signoff; needs_technical_signoff true if any month SLA < 95; agenda_topics = 4 ordered from {partnership_overview, q2_metrics, performance_highlights, q3_initiatives, technical_recovery, commercial_expansion}. Changing agenda did not move the QBR score./exports/churn/train.csv (180 rows), validation.csv (60 rows), candidates.csv (44 rows, no Churn column). 19 features + customer_id (+ Churn target in train/val).tenure, MonthlyCharges, TotalCharges, SupportTickets90d, NPSLast, UsageTrendPct, ActiveSeatRatio (7); categoricals = Contract, PaymentMethod, PaperlessBilling, Partner, Dependents, OnlineSecurity, OnlineBackup, DeviceProtection, TechSupport, StreamingTV, StreamingMovies, InvoicePastDue (12). Target Churn ∈ {Yes, No}.sklearn.pipeline.Pipeline(ColumnTransformer([('num', StandardScaler(), numerics), ('cat', OneHotEncoder(drop='first'), categoricals)]), LogisticRegression(C=1.0, solver='lbfgs')) — all other params default (tol=1e-4, max_iter=100). Converges in ~21 iterations; no convergence workarounds needed.accuracy_band = "90_plus", tenure_coefficient_direction = "negative" (coef ≈ −0.15). feature_count = 19, training_rows = 180, validation_rows = 60.OneHotEncoder(drop='first') is REQUIRED — drop=None yields 91.7% (wrong band). Do not use drop='if_binary'.predicted_churn_probability = predict_proba(...)[:,1] to 3 dp. With default sklearn settings the top-5 candidate probabilities (example set) are stable to 3 dp. Rank candidates by probability desc.deployment_rule_code rule = approve_with_monitoring.outreach_action / reason_code per ranked candidate (BEST-GUESS, principled but unverified): compute each feature's contribution to the log-odds (coef × transformed_value) and label the dominant risk-increasing contributor: InvoicePastDue_Yes→overdue_receivable/collections_followup; low tenure→low_tenure_high_churn/renewal_save; SupportTickets90d→sla_degradation/technical_recovery; UsageTrendPct→usage_decline/technical_recovery; NPSLast→nps_drop/nurture_monitor.
customer_id) — otherwise reasons are assigned to the wrong accounts.cohort_checks: past_due_shortlist_count = count of top-5 with InvoicePastDue=='Yes'; low_tenure_shortlist_count = count of top-5 with tenure < 24; average_probability_top5 = mean of the 5 probabilities (3 dp, computed from full precision not the rounded values).strategic_accounts / enterprise_accounts = counts by segment among the reviewed account list.arr_at_risk = sum of current_arr for accounts whose risk_level is critical or high.open_expansion_pipeline = sum of open expansion opportunity amount (close_date in period, stage not Closed Won/Lost) across the reviewed accounts = sum of per-account expansion_pipeline.net_revenue_exposure = BEST-GUESS: arr_at_risk − open_expansion_pipeline (expansion offsets at-risk ARR). Alternative tested (arr_at_risk alone) gave no score delta because the segment block is scored jointly with the board — verify on a passing board first. Another plausible formula: arr_at_risk + total_overdue − open_expansion_pipeline.RS-2|RS-6|RS-9); pick the one matching the rule you actually applied.receivable_trigger_code=RCP-7 (overdue = 61_90+90_plus), crm_match_code=CM-5 (exact legal_name, subsidiaries unlinked), pipeline_window_code=PW-6 (close_date window), followup_scope_code=FS-4 (all overdue clients). In train_003 the correct option was the 2nd of each triple.risk_accounts[{rank, account_id, risk_score(int), risk_level, primary_action, current_arr(2dp), latest_nps(int), clean_ticket_count(int), overdue_balance(2dp), reason_codes[]}], portfolio_summary{accounts_reviewed, critical_or_high_count, arr_at_risk, collections_count, technical_recovery_count}, model_checks{uses_billing_arr_source: true, tenure_risk_direction: "negative"}, policy_codes.qbr_metrics[{month, revenue(2dp)=recognized_revenue, support_tickets=clean count, sla_compliance_pct(1dp), nps_score}], highlights{average_revenue, peak_revenue_month, peak_revenue, max_sla_month, max_sla_pct, peak_nps_month, peak_nps_score, ticket_trend∈{improving,worsening,flat}}, metric_sources{revenue=billing_snapshot, support_tickets=support_export, sla_compliance=sla_report, nps=nps_survey}, review_plan{review_owner, review_due_date, needs_technical_signoff}, agenda_topics[4].financial_summary{overdue_client_count, overdue_total, linked_followup_count, unlinked_followup_count}, pipeline_summary{won_count, won_revenue, lost_count, open_count, open_pipeline, win_rate_pct, top_open_product_line}, overdue_followups[{customer_name, link_status, account_id|null, overdue_balance, due_date, primary_action=collections_followup}] sorted by customer_name asc, ops_context{hr_headcount, unpaid_claims_total, event_orders, event_revenue}, policy_codes.model_validation{training_rows=180, validation_rows=60, feature_count=19, accuracy_pct=93.3, accuracy_band=90_plus, tenure_coefficient_direction=negative}, risk_ranking[{rank, customer_id, predicted_churn_probability(3dp), outreach_action, reason_code}], cohort_checks{past_due_shortlist_count, low_tenure_shortlist_count, average_probability_top5(3dp)}, model_policy_codes.action_board[{rank, account_id, risk_level, primary_action, current_arr, expansion_pipeline, overdue_balance, next_touch_due_date, reason_codes[]}], segment_summary{strategic_accounts, enterprise_accounts, arr_at_risk, open_expansion_pipeline, net_revenue_exposure}, followup_calendar{collections_followup, technical_recovery, renewal_save, executive_qbr, nurture_monitor}, policy_codes.billing_arr_current or crm_arr instead of the assessment-date billing snapshot.metrics.support_ticket_count instead of the cleaned /tickets count.1_30/31_60 as overdue; only 61_90+90_plus are overdue.region=all to HR/opportunities (returns empty) — use no region filter and sum/filter client-side.OneHotEncoder(drop=None) for the churn model (wrong accuracy band); must be drop='first'.average_probability_top5 from already-rounded probabilities instead of full precision.