Investigate a Spike or Drop in a Metric
Walk through a metric anomaly systematically — what changed, what could explain it, what to check first.
When to use this
When a metric you watch suddenly moved and you need to figure out what happened — without immediately blaming the most visible thing.
The prompt
You are a careful analyst. Your default is "let's see the data" before "I have a theory."
Inputs:
- **The metric** (name, what it measures, how it's computed): [...]
- **The anomaly** — when it happened, magnitude, direction: [...]
- **What was happening at that time** (releases, marketing, seasonality, external events): [...]
- **Recent trends** in this metric over the prior 30 days: [...]
- **Related metrics** that might confirm or contradict the story: [...]
Investigate systematically:
1. **Restate the anomaly precisely** — when did it start, how big is the change relative to typical variance, has it stabilized or is it ongoing?
2. **Confirm it's real** — measurement issues to rule out first: instrumentation change, schema change, bot traffic, time zone or DST, duplicate counting, missing data, definition change.
3. **Generate hypotheses, ranked** — 3–5 candidate causes. For each:
- The hypothesis in one sentence
- The pattern in the data that would CONFIRM it
- The pattern that would FALSIFY it
- Probability (low/medium/high) given what I told you
4. **The one-hour first-pass** — what 3–4 quick checks would you do FIRST to rule out the cheapest hypotheses?
5. **What's missing** — what data would I need to access to be confident in the diagnosis?
6. **The trap to avoid** — what's the obvious-looking explanation that I'm probably wrong about?
Don't pick a cause and defend it. Generate alternatives and let the evidence decide.
What you'll get back
A precise restatement, measurement-check first pass, 3–5 falsifiable hypotheses with probabilities, a quick-check plan, named missing data, and a flagged premature-conclusion trap.
How this is structured in English
Notice the English patterns this prompt uses — they're worth borrowing for your own requests.
- Let the evidence decide Investigator's discipline. Generates hypotheses before picking one; lets data, not narrative, choose. Hard to do when stressed.