Prompt of the day

Churned-user postmortem: exit verbatims + usage timelines to a ranked churn-driver table

A cohort churned last quarter and the room already 'knows why' (price, a competitor, onboarding). Turn the exit surveys and the account usage timelines into an evidence-ranked churn-driver table before that narrative writes the retention roadmap.

You are a product-retention analyst running a churned-user postmortem. Analytics tell you HOW MANY left and WHEN; the exit verbatims tell you WHY. I will paste both. Your job is to turn them into an evidence-ranked account of why this cohort left — a thinking aid whose conclusions I will validate against my own analytics.

Produce:

A) CHURN TIMELINE — for the churned accounts, reconstruct the shared arc from the usage data: signup -> activation -> the value moment (first time the job got done) -> the decay (where frequency/depth inflected down) -> the trigger event -> cancel or lapse. Name THE MOMENT VALUE BROKE: the specific session, week, or event after which engagement fell and did not recover, and the usage signal that marks it.

B) RANKED CHURN-DRIVER TABLE — columns: driver (the reason, in the user's own words where possible), evidence type (verbatim quote / usage signal / both), accounts affected, the receipt (a direct exit-survey quote WITH its response ID, and/or the named usage signal), confidence (high/med/low). Rank by weight of evidence, not by how loud the driver feels. STRICT EVIDENCE RULE — quote-or-it-didn't-happen: every driver row must cite a verbatim quote or a named usage signal I pasted. A suspected driver with neither is written as a row reading "insufficient evidence — needs [WHAT WOULD CONFIRM IT]", never asserted as a cause.

C) LEADING SIGNALS MISSED — the early warning signs visible in the usage timelines BEFORE cancellation (declining login frequency, a core feature abandoned, seats reduced, rising support contacts, export/backup activity), each with the typical lead time before churn — the signals a health score should have caught.

D) JOB-TO-BE-DONE GAP — for each churned segment, the job they hired the product to do and the point at which it stopped getting done, traced to a specific quote. Distinguish 'the product could not do the job' from 'the user never learned it could' — the fixes are opposite.

E) COHORT / SEGMENT PATTERNS — where churn concentrates: by signup cohort, plan, acquisition channel, company size, and activation path. Flag any segment churning materially above the base rate; that is where a fix has leverage.

F) HYPOTHESES + THE ONE EXPERIMENT — the candidate churn causes ranked by the strength of evidence from B-D (not plausibility), separating causes we can act on from ones we cannot. Then name the SINGLE highest-leverage retention experiment to run next: the change, the target segment, the success metric, the guardrail metric, and what result would kill the hypothesis.

Inputs: [PASTE EXIT-SURVEY VERBATIMS: RESPONSE ID, DATE, PLAN, REASON, VERBATIM] · [USAGE TIMELINES / EVENT DATA PER CHURNED ACCOUNT] · [COHORT + SEGMENT DIMENSIONS AVAILABLE] · [THE JOB THE PRODUCT IS HIRED TO DO] · [OVERALL CHURN RATE + BASE RATES IF KNOWN]

Rules: Do not invent churn reasons, quotes, counts, timelines, or usage events — every driver must trace to a verbatim quote or a named usage signal I pasted, and thin evidence is labeled "thin", not upgraded to a cause. Verify churn counts, cohort sizes, and metric definitions against your own analytics before this postmortem enters a retention plan. Anonymize every output: strip real names, emails, and company identifiers — reference accounts by response ID or cohort only. Never paste confidential account, contract, or personally identifying customer data into a consumer AI tool. This synthesizes the evidence; the retention decision stays yours.

Why it works — Churn postmortems die of narrative bias — the room agrees on 'price' or 'a competitor' and the exit data gets read to fit. Forcing every driver in a ranked table to carry a verbatim receipt or a named usage signal (quote-or-it-didn't-happen) makes the loudest theory earn its place, and reconstructing the timeline to the exact moment value broke turns 'they left because of price' into a testable claim about a specific week — so the one experiment you run next is aimed at a cause you can actually see in the data, not the one that was easiest to say out loud.

Catch AI's mistakes by asking what it assumed

Confident-but-wrong is the #1 AI risk at work. Surfacing its hidden assumptions is a 10-second insurance policy on every important output.

After any answer you'll act on, ask: 'What did you assume that I didn't tell you? What would change your answer if it were wrong?'

1. Get the answer.
2. Ask for its assumptions + what would flip them.
3. Verify the load-bearing ones.