Analytics stack

/analytics-stack — The Measurement Backbone

Turn a venture into an instrument you can trust. This is domain 13 (Data & Analytics), paired with /metrics-dashboard: that skill defines what to measure (the KPI framework); this one builds the machine that measures it — the stack, the events that feed it, the experiments that move it, the dashboards that report it, and the governance that keeps it consent-clean and PII-minimal. Industry-agnostic; it reads the active venture and instruments whatever the business already does. Draft/spec only — nothing fires, wires, or reconfigures without Tony's approval.

What this owns vs. delegates (compose, don't duplicate)

This skill OWNS (adds)

Delegated to (pull / run, never re-author)

Analytics stack selection + connect status (web/product/warehouse/BI); the instrumentation plan (event taxonomy + properties, funnel + offer-ladder mapping); the experiment/A-B framework; dashboard assembly + reporting cadence; analytics data governance (consent/PII/retention/DPAs)

North-star + KPI tree + thresholds → /metrics-dashboard · the archetype's money-strip definitions (subscription MRR/NRR/CAC-payback · goods AOV/contribution-margin/repeat/MER · services utilization/realization/rev-per-head · marketplace GMV/take-rate/liquidity) → finance-ops · pipeline/funnel-stage metrics → sales-crm · attribution model + channel CAC + marketing KPI sub-tree → gtm-marketing · chart standards → dataviz · queries/analysis/stat tests/dashboard build → data:* pack · weekly report → /weekly-ops-review · event-capture infra + consent mechanism + PII controls → docs/engineering-backbone.md (starter enforces; venture wires)

Load first

1. founder-profile/PROFILE.md + founder-profile/guardrails.md — governing question, voice, drafts/spec-only, secrets, no-vanity discipline.
2. Active ventures/<slug>/venture-context.md — the archetype (subscription · goods · services · marketplace, or a blend), revenue lines, offer ladder, the core value action, stage, regulatory surface (health/PHI changes the whole governance posture). The archetype selects the funnel shape and the money strip — subscription/SaaS is one archetype, not the default; read it before mapping a single event.
3. /metrics-dashboard output — the north-star + metric tree + thresholds this skill instruments. Do not re-derive the KPI framework here; consume it. If none exists, run /metrics-dashboard first.
4. Where they exist: finance-ops (the money-strip definitions + KPI→P&L bridge), sales-crm (pipeline stages + funnel), gtm-marketing (attribution + acquisition sub-tree + max-CPA).
5. docs/engineering-backbone.md §5 (data/privacy), §7 (observability — consent-gated analytics), §9 (compliance regime) — the governance this skill operationalizes, never contradicts.

Derive every event, metric, and control from these. Never hard-code an industry.

Method (full detail + worked example in references/method.md)

Six parts; for each, name the tool + MCP and flag its connect status. First read the venture's archetype (Load-first #2): it makes the funnel (part 2) and the money strip (part 3) archetype-conditional — never default to the single-sided SaaS AARRR funnel + MRR strip.

1. The measurement stack. Select the layers the venture's stage actually needs — web (GA4), product (PostHog default / Amplitude), warehouse (BigQuery/Snowflake, only when data spans ≥3 systems), BI (Metabase). State each layer's status here (LIVE MCP / REGISTRY-connect / DIRECT-instrument). Ground truth in this environment: the dedicated analytics tools have no live MCP — they connect via SearchMcpRegistry or in-app instrumentation; the live substrate is Supabase MCP (event sink) + Airtable MCP (registers) + the data:*/dataviz/xlsx skills. Don't provision a warehouse or CDP before the volume justifies it (overbuild guard).
2. The instrumentation plan (event taxonomy). Define the tracking plan — the single source of truth. Naming convention: object_action, snake_case, past tense (protocol_executed, checkout_started). For every event: trigger · properties (typed, bounded) · funnel stage · offer-ladder rung · consent tier · PII flag. Map events to the archetype's funnel (subscription/SaaS → AARRR: acquisition → activation → retention → revenue → referral · goods → acquisition → purchase → repeat/reorder · services → lead → qualified → proposal → delivery · marketplace → a two-sided funnel — supply acquisition + activation and demand acquisition + activation → match/first-transaction → liquidity, where each event also carries a side (supply/demand)) and to the offer ladder from /offer-architect. Set identify vs. anonymous, super-properties, and the hard rule: no PII/PHI in event properties (§5/§7).
3. The KPI tree (instrumented). Take /metrics-dashboard's north-star → input drivers → money strip and wire each node to the event(s) that source it, the owning domain, and the query/tool. Leading inputs = product/funnel events; lagging money strip = the archetype's finance-ops definitions (subscription → Engaged MRR/NRR/CAC-payback · goods → AOV/contribution-margin/repeat-rate/MER · services → utilization/realization/revenue-per-head · marketplace → GMV/take-rate/both-sided-retention/liquidity), never defaulting to MRR, + sales-crm pipeline $ + gtm-marketing channel CAC. This is the value-add over the abstract tree: a tree you can actually compute from real data.
4. The experiment / A-B framework. Hypothesis template: "We believe [change] moves [primary metric] from [X] to [Y] because [reason]; we'll know at [significance] within [duration]." One primary metric (OEC) + explicit guardrail metrics; pre-registered MDE, power, sample size, duration (sized via data:statistical-analysis). Decision rule: ship / kill / iterate at pre-set thresholds — no peeking, no post-hoc metric-shopping. Default tool: PostHog experiments (feature flags + A/B). Cap concurrent experiments (focus guard).
5. Dashboards + reporting cadence. Assemble the board /metrics-dashboard spec'd (north-star + ≤7 leading inputs) + a short lagging money strip, laid out to dataviz chart standards, built via data:build-dashboard. Set cadence: real-time ops-health strip · weekly operating board (→ feeds /weekly-ops-review) · monthly money strip / board pack (→ finance-ops). Tool: data:* + dataviz over Supabase at launch; Metabase when a self-serve BI surface is warranted.
6. Data governance. Operationalize the backbone: consent-gated (nothing fires pre-consent; respect DNT — §7), PII/PHI-minimized (UUIDs not emails; bands not raw sensitive values; clinical/PHI never leaves the app DB to analytics — §5/§7/§9), retention schedule + DSAR deletion per data class (§5), DPAs signed with every analytics subprocessor (§9), and no vanity metric as a primary. In health ventures, flag that standard PostHog Cloud / GA4 / Vercel are not HIPAA-eligible — exclude PHI or self-host (§9).

Output contract — a venture's analytics setup

Deliver all five, filled (not blank templates):

1. The stack — the layer→tool→alt→MCP/connect-status→action table (web/product/warehouse/BI + the live substrate), with connect-these flags for anything without a live MCP here.
2. The event-tracking plan — the event taxonomy table (Event | Trigger | Key properties | Funnel stage | Ladder rung | Consent | PII; a marketplace adds a Side column — supply/demand/cross-side), naming convention, identify/super-property rules, and the archetype's funnel + offer-ladder map.
3. The KPI tree (instrumented) — /metrics-dashboard's north-star → inputs → money strip, each node wired to its sourcing event(s), owning domain, and query/tool.
4. The dashboard spec — the board layout (north-star + ≤7 inputs + money strip) to dataviz standards + the reporting-cadence table (View | Audience | Cadence | Source/tool | Feeds).
5. The experiment framework — the hypothesis template + a worked, pre-registered example (primary metric + guardrails + MDE/power/duration + decision rule).

Plus a governance block (consent/PII/retention/DPAs + the regime-specific flag) and any connect-these / needs-Tony flags. Label the whole deliverable [DRAFT/SPEC — nothing instruments, fires, or reconfigures without Tony's approval].

Tools & MCPs

• Stack: Web = GA4 (DIRECT instrument; query via BigQuery export — REGISTRY). Product = PostHog default / Amplitude (both REGISTRY — connect via SearchMcpRegistry + SDK). Warehouse = BigQuery/Snowflake (REGISTRY; defer until ≥3 systems). BI = Metabase (REGISTRY). Error/obs context = Sentry (REGISTRY). None has a live MCP here — flag to connect.
• Live substrate (in-session now): Supabase MCP (event sink + data:sql-queries reads), Airtable MCP (tracking-plan source-of-truth + experiment register), and the skill packs.
• data:* pack + dataviz (the operating engine): data:explore-data (EDA on a new dataset), data:analyze (analysis), data:sql-queries (query Supabase/warehouse), data:build-dashboard (assemble the board), data:create-viz (individual charts), data:statistical-analysis (experiment sizing + significance + power), data:validate-data (tracking-plan QA / data quality), and dataviz (chart standards for every visual). xlsx for offline/board cuts.

Rules

• DRAFT/SPEC ONLY. Produce the plan, schema, and spec; never fire a test event, wire an SDK, change an analytics config, or run a live experiment without Tony's approval. Reversible authoring (the tracking-plan table, the dashboard spec) is done to finished; the activation is gated. (Guardrail.)
• Consent-gated, always. No analytics fires before consent; respect Do-Not-Track; document the taxonomy. (backbone §7.)
• PII/PHI minimization is absolute. No PII/PHI in event properties — UUIDs not emails, bands not raw sensitive values. Clinical/health/financial detail stays in the app DB behind RLS; analytics receives only de-identified behavioral events. Standard PostHog Cloud / GA4 / Vercel are not HIPAA-eligible — in a health venture, exclude PHI or self-host, and treat compliance sign-off as a hard gate. (backbone §5/§7/§9; guardrails.)
• Compose, don't re-derive. Pull the north-star + tree + thresholds from /metrics-dashboard; revenue definitions from finance-ops; funnel/pipeline from sales-crm; attribution + channel CAC from gtm-marketing; chart standards from dataviz. Cite them; run them if missing. Never re-author what a composed skill owns.
• No vanity metric as a primary. Raw pageviews, cumulative totals, follower counts never lead a board or a north-star — convert to a movable, decision-linked rate or demote them.
• Experiments are pre-registered. One primary metric + guardrails, a sizing before launch, a decision rule fixed in advance. No peeking, no metric-shopping. Cap concurrent tests (focus guard); don't build an experimentation platform before you have traffic (overbuild guard).
• Secrets in env only — analytics ingest/API keys are never echoed, logged, or committed; respect the VITE_*/public boundary (ingest keys are public-safe, admin/query keys are not). (backbone §3.)
• Clinical, precise, active voice; banned words out (journey, wellness, holistic, balance, mindful, detox, cleanse, "hack", revolutionary, game-changer, guru, generic hype). Never hard-code an industry.

Composition

• founder-os siblings: /metrics-dashboard (KPI framework — core), finance-ops (money strip + KPI→P&L bridge), sales-crm (pipeline/funnel), gtm-marketing (attribution + acquisition sub-tree), /weekly-ops-review (the weekly board consumer), /offer-architect (the ladder the events map to).
• Skill packs: data:analyze · data:build-dashboard · data:create-viz · data:explore-data · data:sql-queries · data:statistical-analysis · data:validate-data + dataviz (chart standards). xlsx for board cuts.
• MCPs / tools: Supabase MCP (event sink + queries), Airtable MCP (tracking plan + experiment register), and — connect via SearchMcpRegistry — PostHog/Amplitude (product analytics), GA4/BigQuery (web), Metabase (BI), Sentry (error context).
• Backbone: docs/engineering-backbone.md §3 (secrets), §5 (data/privacy), §7 (consent-gated analytics + audit log), §9 (GDPR/CCPA/HIPAA regime).

Examples & evals

• references/method.md — the full six-part method, the stack connect-status table, the event-taxonomy conventions, a complete worked example (Executive Edge analytics setup), and a cross-venture agnosticism check (Ethos hospitality).
• Full deliverable: ventures/executive-edge/outputs/analytics-setup.md.
• evals/evals.md — 3 graded cases (full setup composing not re-deriving; PII/consent/HIPAA governance; experiment rigor + vanity rejection).