Prompt library
Investigate a metric anomaly systematically
Artifact-first ordering reflects the base rates: most metric anomalies are measurement, not reality, and theorizing before ruling that out wastes days. The co-movement test is the cheapest hypothesis killer in analytics and almost nobody runs it systematically.
Last reviewed July 17, 2026
The prompt
Guide an investigation into this anomaly.
The anomaly: {{anomaly}}
Context (systems, recent changes, who noticed): {{context}}
Produce an ordered checklist:
1. Is it real? The measurement-artifact checks first: tracking change, pipeline delay, timezone or DST shift, schema change upstream, bot traffic, one giant account. Each with the exact query or check. Most anomalies die here; check before theorizing.
2. Shape diagnosis: step change, spike, drift, or periodicity break, and what each shape typically implicates (deploys and flags for steps, events and outages for spikes, slow mix-shift for drift).
3. Segmentation ladder: the cuts in order of information value (platform, geography, cohort, plan, entry point), and what a concentrated vs uniform result at each cut would mean.
4. Correlation sweep: which other metrics must have moved if each leading hypothesis is true. Absence of the co-movement kills the hypothesis cheaply.
5. The stop rule: what evidence closes the investigation as explained, and where to log the finding so the next person searching this anomaly finds the answer.
Rule: no cause may be declared while an artifact check remains unrun.Run in idaptOpens a new chat with the prompt prefilled. Nothing sends until you press send.
Fill in the variables
| Variable | What it is | Example |
|---|---|---|
| {{anomaly}} | What moved, when, how much | signups dropped 30 percent last Tuesday and stayed down |
| {{context}} | Recent changes and systems involved | we shipped a new landing page Monday; tracking via 2 tools |