In partnership with
Get what you want from TV advertising
When growth is often measured at the last click, you’re paying to compete for demand that was created somewhere else.
Reach people in the purchase planning phase before your competitors know these customers even exist.
With high-intent Pinterest signals on Performance TV you can reach audiences earlier where they watch the most.
The last edition was about diagnosis. Something broke, and Claude helps you find why. This edition is about something harder and more valuable. Finding the problem before it breaks.
Here is the uncomfortable truth about most growth problems. By the time they show up in your headline numbers, they have already been developing for months. Revenue does not fall off a cliff. It erodes slowly, in cohorts, in segments, in early signals, long before the aggregate number moves enough for anyone to notice. The churn spike you see in Q3 was visible in the Q1 cohort data if anyone had looked. The growth slowdown that hit in December was forecastable in September.
The problem is that nobody looks, because forward-looking analysis is even more time-expensive than diagnosis. Diagnosis at least has a trigger, a number moved, go investigate. Foresight has no trigger. Nothing has visibly broken. So the analysis that would catch the problem early never gets prioritised, and the team operates in a permanently reactive mode, always fixing what has already broken instead of seeing what is coming.
This edition is about using Claude to build an early warning system. A repeatable analysis you run on a schedule, that surfaces the quiet decay before it becomes a visible crisis. Here is how.
What an early warning system actually looks at
Aggregate numbers are lagging indicators. They tell you what already happened. An early warning system looks at leading indicators, the things that move before the headline number does. There are four worth watching, and Claude can analyse all of them.
The first is the cohort quality trend. Not whether customers are retaining, but whether each new cohort of customers is retaining worse than the one before it. A business can have a stable overall retention number while every new cohort is quietly weaker, because the strong old cohorts are propping up the average. When the old cohorts age out, the number falls, seemingly suddenly. Cohort quality trend sees it months early.
The second is engagement decay inside the active base. Customers do not churn the day they stop paying. They churn weeks or months after they stop engaging. A drop in usage, logins, or activity inside your still-paying customer base is a leading indicator of churn that has not happened yet but will.
The third is the shape of your funnel over time, not its level. Your conversion rate might be stable, but if the conversion is increasingly dependent on one channel, or one segment, or one type of customer, the funnel is getting more fragile even though the headline number looks fine. Fragility is a leading indicator of a future drop.
The fourth is acquisition quality drift. Whether the customers you are bringing in this month are as good, by every measurable signal, as the ones you brought in six months ago. Acquisition quality almost always degrades slowly as a company scales channels, and it shows up in revenue much later than it shows up in the data.
An early warning system is just a regular, scheduled analysis of these four things. Claude makes it cheap enough to actually do.
The workflow
Step one: decide the cadence. Foresight analysis is not a one-off; it is a habit. Monthly is right for most teams. Put it on the calendar as a recurring block, the same way you would a financial close.
Step two: Prepare a consistent export. Because you are running this every month, the value compounds when the data is in the same shape every time. Set up a standard export. Cohort data by acquisition month. Engagement data for the active base. Funnel breakdown by channel and segment. Acquisition data with whatever quality signals you have. The first month takes effort to assemble. Every month after, it is a repeat.
Step three, give Claude the current export and, critically, the previous months. Foresight is about trends. Claude cannot see decay from one month of data. It needs the sequence. Each month, you give it the new data plus the prior months, and you ask it not what the numbers are, but which direction they are moving and how fast.
Step four, instruct Claude to look for divergence and deceleration, not absolute levels. The early warning signal is never a number being low. It is a number trending in the wrong direction while the headline number still looks fine. That divergence, healthy aggregate, deteriorating leading indicator, is the entire thing you are hunting for.
Step five: make it output a watch list, ranked by urgency. The deliverable is not a report. It is a short, ranked list of the things that are quietly deteriorating, how fast, and how long until each one becomes visible in the headline numbers if nothing changes.
The operational prompt
Here is the complete early warning prompt. Run it monthly, attach the current export and the prior months, and fill in the context.
You are a senior growth analyst running a monthly early warning
review. Your job is NOT to report what the numbers are. Your job
is to find problems that are developing but have not yet shown up
in headline metrics.
CRITICAL RULES:
- Work only from the data I provide across all the months
attached. Do not estimate, recall, or invent any number.
- Focus on TREND and DIRECTION, not absolute levels. A number
being low is not the signal. A number moving the wrong way
while headline metrics still look healthy IS the signal.
- If the data is insufficient to assess a trend, say so. Do
not manufacture a trend from too few data points.
CONTEXT:
- Business: [what you do, business model, customer]
- Headline metrics and their current state: [e.g. revenue,
overall retention, overall conversion, and what each is doing]
- The data attached: [describe each file and the months covered]
YOUR TASK, in this order:
1. COHORT QUALITY TREND. Compare each acquisition cohort to the
earlier ones. Is each new cohort retaining and monetising as
well as previous cohorts, better, or worse? Quantify the
direction and the rate. Flag if the overall retention number
is being propped up by older cohorts.
2. ENGAGEMENT DECAY. Within the active, still-paying base, is
engagement (usage, logins, activity) trending down in any
segment? Identify any group whose engagement is falling even
though they have not yet churned.
3. FUNNEL FRAGILITY. Is the funnel becoming more dependent on a
single channel, segment, or customer type over time, even if
the overall conversion rate looks stable? Concentration is
fragility. Flag it.
4. ACQUISITION QUALITY DRIFT. Are customers acquired in recent
months as strong, on every measurable signal, as customers
acquired earlier? Identify any decline in incoming customer
quality.
5. THE WATCH LIST. Produce a ranked list of everything that is
quietly deteriorating. For each item: what is decaying, the
evidence in the data, how fast it is moving, and your best
estimate of how long until it becomes visible in a headline
metric if nothing changes. Rank by urgency.
6. THE SINGLE PRIORITY. Of everything on the watch list, name
the one thing I should act on first this month, and why.
If the data across the attached months is insufficient to assess
any of these trends reliably, tell me which one and what I should
export to fix it.What changes when you run this
A team that runs diagnostic analysis is reactive. It is good at fixing things, but it only ever fixes things that have already broken, which means it is permanently dealing with problems at their most expensive and most visible stage.
A team that adds the early warning system changes its relationship to its own numbers. It starts seeing problems while they are small. The cohort decay gets caught in month two instead of month eight. The engagement drop in a customer segment gets caught before the churn, while there is still time to intervene. The funnel fragility gets noticed before the channel it depends on has its bad month.
The financial difference between catching a problem early and catching it late is enormous, and it is almost entirely invisible, because a problem caught early never becomes a number anyone sees. The team just quietly has fewer crises, and nobody can point to a crisis that did not happen.
This is the highest form of data work. Not explaining what happened. Seeing what is about to happen, while there is still time to change it.
These two editions together give you both halves. Diagnosis, for when something has already moved, and you need the real cause, fast. Foresight runs on a monthly schedule, for catching the next problem while it is still small. A growth team that does both is no longer at the mercy of its own dashboard. It understands its numbers, backwards and forwards, and it acts on understanding instead of guessing.
Set the monthly block on your calendar. Build the standard export once. Run the prompt. The first month, it will tell you something about your business that you did not know. Every month after, it will be the cheapest insurance your growth team has.
See you at the next edition, Arindam