Product analytics instrumentation review

Product Analytics Instrumentation Review

Produce an Analytics Event, Identity, and Metric Contract that makes product decisions reproducible without turning surveillance, dashboard convenience, or client events into false authority.

Atomic boundary

Own decision-to-signal mapping, event/property/metric semantics, identity/session, consent/privacy, SDK/server collection ports, delivery/quality, warehouse/join contracts, QA, dashboards, exposure measurement, backfill, and drift. Do not own whole product strategy, provider billing truth, marketing spend attribution control plane, experiment decisions, or feedback prioritization. Service/runtime telemetry, health, SLOs, alerting, and operator diagnostics are operational observability, even when the same pipeline also carries product events. Keep their purposes, schemas, access, retention, and authorities separate.

Read references/data-quality-and-metric-layer.md when the request includes warehouse or semantic metrics, conflicting dashboards, dataset trust states, quality monitoring, backfill, certified metrics, access, or cross-source reconciliation. Product measurement quality belongs here; generic data platform access, pipeline engineering, AI dataset assurance, billing truth, and incident implementation remain with their canonical engineering, payment, and applicable binding Skills owners.

Use the shared artifact envelope only when composing with repository product artifacts. For a narrow audit, include only the contract surfaces needed by the declared decisions. A disabled SDK initializes nowhere, sends nothing, and collects no identifier.

Workflow

1. List decisions and machine actions first. For each, name the outcome, mechanism, segmentation, countermetrics, latency/freshness, confidence, and authority needed. Reject events with no declared decision consumer.
2. Read references/product-analytics-instrumentation-patterns.md. Map the canonical journey and state transitions, including pending, committed, failed, recovered, suppressed, reverted, and support-corrected outcomes. Load references/data-quality-and-metric-layer.md for warehouse, semantic metric, quality, trust-state, reconciliation, or backfill work.
3. Define an event namespace and semantic version; required/optional properties, types/enums/units, timestamps, IDs, causality, idempotency, actor/source, privacy class, retention, owner, and deprecation/migration.
4. Separate client intent/UI, server/business authority, provider/payment, experiment exposure, marketing touch, support case, quality/error, and derived metric inputs. Critical truth is server/provider authoritative.
5. Define anonymous/device/user/account/organization identities, login/logout, guest upgrade, merge/split, deletion, shared devices, cross-platform, pseudonymization, and no-consent/child/territory modes.
6. Specify consent-aware SDK ports, lazy initialization, offline/batch/retry, sampling, late/out-of-order/duplicate handling, bot/internal traffic, clock/timezone, data residency, deletion/export, and zero-cost dormant state.
7. Define any consumed experiment exposure, attribution, billing, and support joins without taking ownership of those domains; specify metric definition ownership, dashboard freshness, dimensions, and source lineage.
8. Build representative fixtures for material event/version/platform/state combinations; add contract validation, golden journeys, quality checks, correction/backfill policy, and release gates proportional to failure risk.

Source verification

Retrieve current analytics/ads SDK, platform privacy manifest, consent, child and regional privacy, ATT/device identifier, cookie/storage, data-residency, deletion/export, and provider quota/retention authority. A vendor default is never the product's consent or retention policy.

When not to use

• Use app-design-blueprint, game-design-blueprint, or product-lifecycle-architect when the primary artifact is product behavior or a cross-domain delivery program, not measurement implementation.
• Use marketing-automation-blueprint for spend, channel attribution decisions, creative automation, and budget/shutdown control.
• Use the owning product or experiment workflow to choose hypotheses, variants, exposure, and promotion; analytics records the agreed assignment and outcome.
• Use product-feedback-learning-loop for qualitative feedback/review ingestion, evidence clusters, support routing, and product close-loop. Use review-solicitation-policy for public review request eligibility and state.
• Use payment-platform-readiness for payment/entitlement/settlement authority; analytics only consumes its signed/authoritative projection.
• Use operational-observability-review for service logs, traces, health, SLOs, alerts, runtime diagnostics, and operator action. Use data-quality-observability-review when generic dataset or pipeline reliability is the independently accepted artifact.

Guardrails

• Do not instrument everything or collect data “just in case.” Minimize by declared decision and retention need.
• No client event, dashboard, or model inference may grant entitlement, settle money, enforce policy, or overwrite operational truth.
• Do not merge identities without explicit rules, consent/authority, reversibility, and deletion semantics.
• Never silently change an event or metric meaning. Version, dual-write/read, backfill or annotate discontinuity, and migrate consumers.
• Dashboard freshness/coverage and model confidence are not product success; retain value, quality, trust, fairness, privacy, and support countermetrics.

Output contract

Return one typed Analytics Event, Identity, and Metric Contract with:

1. decision-to-signal/countermetric map and authority classification;
2. event/property schema registry, semantic versions, lineage, owners, privacy, retention, and deprecation/migration;
3. identity/session/account merge/split/logout/delete model;
4. client/server/provider/experiment/marketing/support collection and join map;
5. consent/age/territory/platform modes, SDK ports, dormant/offline/retry, sampling, dedupe, and correction;
6. canonical metric definitions and dashboard/consumer contracts;
7. fixtures, golden journeys, data-quality checks, correction/backfill, release gate where warranted, and live-readback plan;
8. explicit sibling handoffs where the measurement contract consumes another domain's authority.

Complete only when every metric traces to versioned events and authority, every event has a decision/owner/privacy lifecycle, and synthetic plus live journeys prove exact-candidate data quality.