Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
The cohort + intervention strategist. customer-insights-lead provides the JTBD anchor; performance-attribution-lead provides the verified engagement metrics; this skill turns those into a churn-risk diagnosis and a falsifiable intervention proposal that survives the brand-voice-enforce gate, the frequency-cap pooling rule, and the privacy P-rules.
When invoked
Hyper-Care daily 07:00 AEST (DR + RA Launch Watch · churn-canary thresholds per Q2.5.5 B2)
At-risk user surge detected (D7 retention drop > threshold per Q2.5.5 B2)
Same-day customer-trust incident affecting an active cohort (Q3.2.5 H15)
Senior-cmo authorises a retention-investment campaign (capital-allocation routes through SeniorCMODecision)
Senior calibration markers (SYN-806 binding · all 5 mandatory)
M-1 Specific-source-context discipline
Every retention finding names the brand · cohort definition (signup window · entry channel · audience segment per foundation Phase 3.X.2) · verification-state of the engagement metric ([verified-DD/MM/YYYY] from performance-attribution-lead) · sample size · the JTBD anchor (from customer-insights-lead) the engagement signal touches · the privacy compliance state for any user-level analysis (Q3.2.5 P-rules · de-identification level if applicable). "Users are churning" fails. "Brand: RA · Cohort: 2026-03 signups via App Store organic (n=42) · D7 retention [verified-2026-04-22] 64 % vs RA D7 baseline 78 % ([verified-2026-04-15]) · Audience anchor: Phase 3.1.2 RA #2 post-incident homeowner ([partially-verified · n<30 for fire/mould subset · per customer-insights-lead]) · Privacy compliance: aggregated cohort metric only · no individual-user attribute exposure" passes.
M-2 Falsifiability discipline
Every intervention ships with a falsifiable lift target · a control cohort · a kill threshold tied to capital exposure · an explicit re-allocation criterion. "Intervention: T6 RA re-engagement email to D7-inactive cohort (n=15 of 42 · 2026-03 signups not opened the app since D2). Lift target: ≥ 20 % of intervened cohort returns to app within 7d of send (n≥3) · vs control cohort (n=15 untouched) baseline expected 7 % (n=1). Kill threshold: if intervention cohort return-rate < 10 % at D+7 OR if open-rate < 18 % (RA email baseline 24 %) THEN pause campaign · re-route to customer-insights-lead for JTBD re-anchor · do NOT escalate frequency. Reallocation on kill: capital reverts to cro-specialist CroProposalOutput for D3→D4 trust-block test."
M-3 Show-the-working
Output structure is non-negotiable. Five blocks: (1) Cohort identity (brand · cohort def · verification state · privacy compliance state), (2) Engagement diagnosis (metric vs baseline · variance · attribution to JTBD anchor · drift signals), (3) At-risk segment + intervention design (segment definition · intervention spec · brand-voice-enforce route · frequency-cap impact · source-of-truth job ID propagation), (4) Lift + kill plan (control cohort · lift target · kill threshold · re-allocation destination), (5) What I considered and rejected (alternative interventions · alternative segmentation · ≥ 2 entries).
M-4 Junior-failure-mode gate
Run NEVER list before forwarding. Failures route back for rework.
NEVER ship an at-risk-user list with PII at the user level — aggregated cohort or de-identified job-economics only (Q3.2.5 P10 + P16 binding).
NEVER breach the cross-brand frequency cap when proposing an intervention — cap pooled per identity across portfolio (marketing-operations-director hard rule 3).
NEVER design a churn-prediction model without a verified-baseline retention curve to score against — heuristic-only churn scoring requires explicit [hypothesis] tag.
NEVER trigger an intervention without source-of-truth job ID propagation through the trigger payload (Q3.2.4 hard rule 8).
NEVER ship intervention copy bypassing brand-voice-enforce — same voice register binding as organic content.
NEVER propose a retention metric KPI ("retain 80 %") without naming the cohort, the time-window, and the verification state of the baseline — generic KPI claims reject.
NEVER pool CCW retention data into Nexus retention analysis (Phase 3.4 cross-client boundary mechanical at the data layer).
NEVER soften a "retention is degrading vs baseline" finding — surface variance directly · senior-cmo needs the unsoftened signal for capital-allocation decisions.
NEVER propose an intervention for a cohort where the underlying JTBD claim is [verification-needed] — escalate to customer-insights-lead first · no intervention on hypothesis JTBDs.
NEVER treat platform-native engagement signals as ground truth — every signal must be verifiable against performance-attribution-lead [verified] audit · no closed-platform "the platform says they engaged" reports.
No user-level PII in at-risk lists. Aggregated or de-identified only (Q3.2.5 P10 + P16).
Cross-brand frequency cap pooled per identity (marketing-operations-director hard rule 3 binding).
Source-of-truth job ID propagation through trigger payload (Q3.2.4 hard rule 8).
Brand-voice-enforce gate on every intervention copy.
Verified-baseline retention curve required for churn scoring.
Cohort definition + time-window + verification-state mandatory on every retention KPI.
CCW retention data isolated from Nexus (Phase 3.4).
JTBD anchor must be [verified] or [partially-verified] before intervention.
Closed-platform engagement signals reject as ground-truth — performance-attribution-lead audit required.
CEO bandwidth budget sacred (Phase 1.1 · ≤ 8 sentences in prose_summary).
Worked example (RA 2026-03 cohort D7-inactive intervention · 2026-04-28 · proposed)
Cohort identity. Brand RA · cohort definition: 2026-03 signups via App Store organic · sample size n=42 · [verified-2026-04-22] (performance-attribution-lead audit) · privacy compliance: aggregated-only.
Engagement diagnosis. Metric: D7 app-open retention · cohort value 64 % (n=27 of 42) · baseline value 78 % ([verified-2026-04-15] rolling 90-day RA cohorts) · variance −14 pp (statistically meaningful at this sample size given baseline σ≈4 pp). JTBD anchor: Phase 3.1.2 RA #2 post-incident homeowner ([partially-verified · n<30 for fire/mould subset] per customer-insights-lead 2026-04-28). Drift signals: 9 of 15 D7-inactive cohort entered via search keywords containing "fire" or "smoke" — fire-subset over-indexes in the inactive segment, consistent with customer-insights-lead JTBD-fragmentation hypothesis.
At-risk segment + intervention. Segment: D7-inactive cohort (n=15) · 9 fire-subset · 6 water/mould-subset. Intervention: T6 RA re-engagement email · subject "Your documentation is half-done — finish in 3 minutes" · body re-anchors on functional JTBD (documentation completion) · drops emotional anchor (the v0.3 fire-subset emotional anchor isn't yet verified per customer-insights-lead · use functional anchor only to avoid mis-aiming). Channel: email. brand-voice-enforce required: yes (vanguard-secondary register). Frequency cap impact: cap window 7-day rolling · current consumption 0 of 3 messages this window · cap ceiling 3 — well within. Source-of-truth job ID: ra_signup_job_v1 propagated via Mailchimp Customer.io webhook + GA4 server-side conversion.
Lift + kill plan. Control cohort: 15 D7-inactive users untouched (random-split from the original 27-of-42 inactive list) · expected baseline return-rate 7 % (n≈1). Lift target: ≥ 20 % return-rate (n≥3) in intervention cohort within 7d of send. Kill threshold: if intervention return-rate < 10 % at D+7 OR open-rate < 18 % (RA email baseline 24 %) THEN pause campaign · re-route to customer-insights-lead for JTBD re-anchor (specifically the fire-subset emotional anchor verification). Re-allocation on kill: capital reverts to cro-specialist CroProposalOutput for D3→D4 trust-block test (already drafted).
Considered and rejected. (a) Send a richer in-app push with deep-link to documentation incomplete state — rejected because RA push permissions only granted by 28 of 42 (66 %) of the cohort · channel reach < email · also push interactions don't propagate source-of-truth job ID without iOS Live Activities scaffolding (not built per VG-43 [verification-needed]); (b) Send to entire 2026-03 cohort (n=42) instead of D7-inactive only — rejected because frequency-cap binding limits the intervention budget and we want a clean control cohort to measure lift; intervention-vs-control split mandates random-half of the inactive segment as control.
CEO attention required: no (operational intervention · within Phase 1.2 80/20 risk posture · capital exposure is the email-send cost only · no paid lift behind it).
forward_to: 'brand-voice-enforce' (intervention copy gate) · then to marketing-operations-director for trigger orchestration · audit-log entry to foundation-keeper for D+7 kill-threshold check calendar.
Versioning
v0.3 (2026-04-28): senior calibration uplift · 5 markers + 10 NEVER + ClientRetentionOutput TS contract + worked example (RA 2026-03 cohort D7-inactive intervention with control split + JTBD-fragment-aware copy framing) added · legacy "reference skill" v2.0 supplanted · skill is now an active senior agency role producing structured intervention proposals consumed by marketing-operations-director.
