🇰🇷 한국어 README
Data analytics engine built on a 5-layer framework (Question · Data · Analysis · Interpret · Deliver) with a 6-domain router and a 16-trap defense library. LIGHT · DEEP · TURBO 3 modes.
- Claude Cowork or Claude Code environment
- Works best when paired with
research-frame,metric-tracker, andfinancial-modelfor domain handoffs
Data analysis fails at the extremes — the question (wrong thing measured) and the interpretation (wrong thing concluded) — not at execution. This skill keeps humans in charge of those endpoints and delegates the middle (query, aggregation, computation) to AI, while a 16-trap checklist audits every layer. It covers product analytics, marketing analytics, unit economics, A/B experiments, root-cause investigations, and metric governance in one coherent pipeline.
Use this skill when you have a real data question ("why did MAU drop 2pp?", "design an A/B test for this feature", "diagnose this conversion rate"). The skill auto-routes to one of 6 domain modules (M1 Product, M2 Marketing, M3 Finance/UE, M4 Experiment, M5 RCA, M6 Governance) and one of 3 modes (LIGHT for quick diagnosis, DEEP for full pipeline, TURBO for parallel multi-domain analysis). Distinct from metric-tracker (periodic KPI tracking) and financial-model (forward-looking projections) — this one handles diagnostic and causal questions on existing data.
| Scenario | Prompt | What Happens |
|---|---|---|
| Metric drop investigation | "MAU가 2%p 떨어졌어 왜?" |
DEEP · M5 RCA — L1 question framing → L2 definition check → L3 segment/cohort/funnel decomposition → L4 Simpson/definition traps → L5 1-page RCA memo |
| A/B test design | "신기능 A/B 설계해줘" |
DEEP · M4 — L1 hypothesis + MDE → L3 power analysis + sample size → L4 p-hacking/SRM/leakage checklist → L5 readout template |
| Quick number triage | "이 숫자 빠르게 진단: 3.2%→2.8%" |
LIGHT — L1 question in 1 line + L4 top 2-3 traps, confidence capped at 50 |
| Multi-domain quarterly deep dive | "이커머스 분기 deep dive 터보로" |
TURBO — parallel agents for Product + Marketing + Finance, sequential L1 gate + L4 cross-check + L5 integrated memo |
- 5-Layer Framework — L1 Question (human) → L2 Data (partial AI) → L3 Analysis (strong AI) → L4 Interpret (human) → L5 Deliver (partial AI). Human owns the endpoints; AI owns the middle.
- 6 Domain Router — M1 Product · M2 Marketing · M3 Finance/UE · M4 A/B Experiment · M5 RCA · M6 Governance. Auto-selected by keyword and intent.
- 16-Trap Library — Statistical (Simpson, Survivorship, p-hacking, Goodhart, Base rate), Cognitive (Confirmation, Anchoring, Storytelling), Infrastructure (Definition mismatch, Timezone mismatch, Leakage, Sampling bias), Visualization (Truncated axis, Average-only), Org/AI (Conclusion-first, Text-to-SQL errors).
- L1 & L4 Gates — Mandatory 5-point checklists at question-design and interpretation layers. No L2 entry without L1. No L5 delivery without L4.
- AI Delegation Boundaries — Text-to-SQL is 64.5% accurate without a Semantic Layer, 95%+ with one. The skill specifies where AI may run, where it must stop.
- Confidence Quantification — UP-aligned: High 90 / Normal 70 / Low 50 / Unknown 30, with a one-line justification. Single-source data caps confidence at 70.
- 3 Modes — LIGHT (1-page diagnosis, cap 50), DEEP (full pipeline), TURBO (parallel domain agents).
- research-frame — Qualitative research upstream. When the question is "what do users think?" not "what does the data show?"
- metric-tracker — Periodic KPI/OKR tracking. Data-analytics handles why a metric moved; metric-tracker handles whether its on target.
- financial-model — Forward projections (VC-grade revenue models, scenarios). Data-analytics looks backward; financial-model looks forward.
- biz-skill — Strategic pattern matching. Consume data-analytics findings as evidence for strategy decisions.
- fact-checker — Numerical verification. Use after data-analytics produces claims for external delivery.
git clone https://github.com/jasonnamii/data-analytics.git ~/.claude/skills/data-analyticscd ~/.claude/skills/data-analytics && git pullSkills placed in ~/.claude/skills/ are automatically available in Claude Code and Cowork sessions.
This is one of 40+ custom skills. See the full catalog: github.com/jasonnamii/cowork-skills
MIT License — feel free to use, modify, and share.