analytics-process

name	analytics-process
description	Use when planning, executing, reviewing, or calibrating a product analysis, or when asked a product data question that should become one (the entry runbook routes by product — Win and Serve today). For the data facts themselves, use the product's knowledge skill (win-analytics-knowledge or serve-analytics-knowledge).

Analytics process

How a product analysis is run, from a fuzzy question to a reviewed result. For the data facts themselves (which table, which metric, which join), use the product's knowledge skill (win-analytics-knowledge, serve-analytics-knowledge) — this skill is about how to work, not what is true about the data.

The senior-analyst loop

An index, not the specification: each step's operational detail lives in its reference doc, and the stage/actor/gate topology lives in references/pipeline.md, the single description of the pipeline flow. Two legibility conventions are always on — announce each step transition with a one-line stage banner, and each cleared checkpoint with a gate receipt; formats and stock why-clauses live in pipeline.md ("Legibility").

1. Clarify and scope (framing). Run the framing routine (references/framing.md) in your own context — it converses with the user. No execution code until the human approves the framing.
2. Find sources. Resolve every concept through the knowledge skill, and verify named tables/columns/events against the live catalog — docs drift.
3. Produce the brief. The framing output is a structured brief per references/brief-schema.md, the framing→execution contract.
4. Execute. Build the working set once, then slice every cut from it in pandas — the discipline is in references/methodology.md; the product's defaults, cohort tables, and working-set column caveats are in the product knowledge skill's methodology_defaults.md.
5. Review. After the results checkpoint, route the executed analysis through the two adversarial reviewer agents; the gate and dispatch rules are in pipeline.md.
6. Close the loop. Run the calibration pass (references/calibration.md), ending with the candidates-ledger read-back it requires. Pruning happens via consolidation mode, detailed in the same doc.

Reusable analysis patterns

Common work that should not be reinvented each run:

• Build-once-slice-many. One consolidated per-user cohort × engagement working set via the product's committed builder, carrying the standard slice dimensions so re-cuts are free pandas groupbys. Builder and column caveats: the product knowledge skill's methodology_defaults.md.
• Funnel analysis. Aggregate the raw event stream to user_id × funnel step; report "engaged beyond account creation" (≥2 distinct events) alongside raw any-evidence. Event families come from the product knowledge skill (Win: engagement.md; Serve: sources.md).
• Rate decomposition. Don't pool heterogeneous funnel stages — name the cohort (onboarded / activated / full registered) so a rate isn't dominated by never-active registrants. See the default-cohorts table in the product knowledge skill's methodology_defaults.md.
• Retention / reverse-retention curves. Pair any closed-cohort curve with an open fixed-denominator view, then split open into new vs returning. The closed curve decays by construction; the open/returning view answers whether the base is more engaged. Full mechanism in the knowledge skill's gotchas.md.
• Wilson intervals + binning. Report Wilson 95% CIs where they inform the read (small N, or a claimed difference) rather than on every cut; flag N<30 bins; pre-register bins in the brief. See methodology.md.

Reference docs

• references/framing.md — the framing routine (persona, question set, pre-brief verification) the orchestrator runs in step 1.
• references/pipeline.md — stages, agents, handoff contracts (descriptive, not active).
• references/methodology.md — product-agnostic discipline: scoping checklist, folder patterns, query patterns, binning, verification protocol. Product defaults (scoping decisions, cohorts, working-set builder) live in the product knowledge skill's methodology_defaults.md.
• references/brief-schema.md — the analysis-brief YAML contract + amend-vs-reframe rule.
• references/calibration.md — the calibration-log convention, the candidates ledger, the ledger read-back, consolidation mode, and the over-calibration / bloat cautions.
• references/event-lifecycle-assets.md — the omni event-lifecycle assets (provenance CSV, health log, gp-meta): cross-product source material consumed by the framing version-continuity check and every product's knowledge skill.