An analytics memo should carry the metric contract version, source snapshot, denominator ledger, segment table, data-quality checks, uncertainty, decision owner, and proposed next observation. Prompt the model to draft claims only from locked figures and to name what cannot yet be inferred. A release gate should fail if a segment total does not reconcile, an immature cohort is presented as final, or a hypothesis becomes a causal claim. If a product change proceeds, define its success measure before exposure and monitor instrumentation health alongside outcome metrics.
Make this comfortable
Analytics prompts: turn a metric packet into a bounded decision
Operational case
Meridian's memo states that 2,820 of 4,700 eligible starting users completed within 24 hours. It reports 50% mobile and 70% desktop, with a 20-point gap. The decision is to inspect mobile errors and event delivery, then design a controlled test for a shorter form if the measurement holds. The memo does not say the new form will improve conversion. An owner records the query snapshot and schedules a recheck after the last cohort matures.
Claim: 2,820 / 4,700 = 60% within 24 hours
Segments: mobile 50%; desktop 70%; gap 20 points
Decision: inspect errors and event delivery
Hypothesis: shorter form may help; unproven
Gate: totals, cohort maturity, and owner verifiedPerformance and review cost
Reconciling C claims with an indexed result ledger is O(C) lookups plus review time; monitoring N new events is O(N) processing. A concise decision memo reduces repeated analytical work when it stores the contract and query version. Do not optimize for a persuasive narrative at the expense of reproducibility or a clear unknown state.
Common Mistakes
- Do not publish a causal headline from an observational funnel.
- Do not omit a query or metric version from the handoff.
- Do not set a success target after viewing the intervention result.
Connected lessons
- Production prompt engineering
- Prompt Engineering
- Report prompts: reconcile every claim with the packet
- Evaluation sets: measure the failure cases that matter
- Analytics prompts: define the metric and event contract
- Analytics prompts: reconcile identities, duplicates, and late arrivals
- Analytics prompts: enforce funnel eligibility and event order
- Analytics prompts: pin cohort windows and time zones
- Analytics prompts: review read-only queries and totals
- Analytics prompts: explain segment gaps without inventing causes
- Project: reconcile Meridian's signup funnel
- Product analytics prompt decisions
Continue with: Experiment prompts: gate rollout on a reproducible packet.
prompt engineering
product analytics