Know whether you can measure it, before you ship it.
Every growth lead knows the feeling. You ship the thing, you wait three weeks, and then you find out the one event you needed to judge it was never added. There is no way to go back and collect it.
Four steps, and the useful one happens before launch.
You find out a required signal is missing while you can still add it, and Skene names which one.
Describe what you are shipping
Verify you can measure it
Ready to launch, or not yet
Check it afterwards
Activation: first order within seven days of signup
Draft reviewed. Gathering evidence from code and arriving events.
- Plan
- Verify
- Launch ready
- Check
It names the signal, not a score.
Three evaluations, each described in a sentence and then verified against the code and the events actually arriving. None of the three reached launch ready, and the screen says why rather than rounding it up. The last one needs a checkout_abandoned signal that is not being recorded anywhere, so the metric it is built on cannot be computed at all.
| Evaluation | Status check | Metric | Confirmed |
|---|---|---|---|
| Activation: first order within seven days of signup | verifying | First-week first-order activation rate | 0 / 10 |
| Flag customers that have gone quiet for thirty days | verifying | Quiet customers resuming orders within thirty days | 0 / 13 |
| Recover an abandoned checkout by email | verifying | Abandoned checkout recovery rate within seven days | 0 / 15 |
public.customers, public.orders and public.invoices, and an abandoned checkout writes no row in any of them, so the denominator cannot be built at all. The evaluation names the signal it is missing rather than returning a number that would look fine.Two data points are two data points.
Every verdict shows the previous number and reads up to ten prior runs, and says so when that history is too short to trust.
When the Check runs, it does not just hand you a number. It reads the hypothesis you wrote, what the formula returned this time and last time, and up to the last ten runs with their own verdicts, then says what it thinks happened.
A short history is treated as weak evidence and labelled that way. If your plan has no target, it does not invent a pass or fail: it judges movement and the quality of the evidence behind it.
A verdict you can take to a review
The number, the previous number, the change, and every operand that went into it. Not a score with nothing underneath.
Where the evidence is thin, it says the evidence is thin.
Nobody is asked to write event names.
What reaches engineering is a named signal and the place it is missing, or nothing at all when the tracking is already there.
You describe the change in the language you would use in a planning meeting. Skene works out which signals that implies and whether they exist. Engineering gets a specific missing event or field, or nothing at all if the tracking is already there.
That is the same split as the release check, moved one step earlier: you decide what has to be answerable, Skene turns it into something checkable.