FLAGSHIP SYSTEM — IN USE DAILY · UNDER ACTIVE BUILD

Analyst Vault.

An AI-native analytics workflow — eight skills, one orchestrator.

CLAUDE CODE · MARKDOWN-FIRST · REPRODUCIBLE BY RULE

STATUS — DAILY DRIVER BUILT — 2026 THIS SITE WAS MADE INSIDE IT
00 — WHY IT EXISTS

Analysis quality is mostly process quality: the questions you ask before writing code, the checks you run before trusting a number, the honesty of the final writeup. AI assistants are powerful but unstructured — they'll happily skip straight to SQL. The Vault encodes the process itself, so every analysis follows the same spine: context before queries, validation before claims, uncertainty stated out loud.

01 — ARCHITECTURE
BUSINESS QUESTION GLOBAL RULES — CLAUDE.MD: DOCUMENT EVERYTHING · REPRODUCIBLE BY DEFAULT · SURFACE UNCERTAINTY · SCOPE DISCIPLINE ORCHESTRATOR — ANALYSIS-WORKFLOW CONTEXT — BUSINESS FRAMING EXPLORE — PROFILE & CLEAN METHODS — SEGMENT & DERIVE STATISTICS — VALIDATE SQL — QUERY CRAFT VIZ + REPORT — DELIVER SKILLS/ — LEARNED KNOWLEDGE PRINCIPLES · ANTI-PATTERNS · PLAYBOOKS (GROWS PER PROJECT) DELIVERABLE — V1.0 FIG. V-01 — VAULT / FULL SCHEMATIC · EVERY ARROW IS A PHASE GATE

Three layers. Global rules (a CLAUDE.md read at every session start) bind all output — documentation, reproducibility, honesty about limits. The orchestrator decides which phase comes next and carries context between them. The skills hold the craft: each one is a written playbook with methods, anti-patterns and mandatory sub-steps — the kind of discipline a good senior would enforce in review, encoded so it happens every time.

02 — THE EIGHT SKILLS
ORCHESTRATION analysis-workflow — sequences the skills, manages phase transitions, builds shared context progressively.
CONTEXT business-context-exploration — goals, success metrics, hypotheses and data requirements before any technical work.
TECHNICAL data-exploration — profile structure and types, find quality issues and anomalies; cleaning is a built-in sub-phase, not an afterthought.
TECHNICAL analytical-methods — segmentation, aggregation, feature creation: clean data into actionable findings.
TECHNICAL statistical-analysis — hypothesis tests, significance, regression; invoked whenever a finding needs mathematical backing.
TECHNICAL sql-query-writing — query construction and optimization across DuckDB, PostgreSQL, BigQuery and Snowflake dialects.
OUTPUT data-viz-dashboarding — chart selection and dashboard design, with a mandatory prototyping step before any code is written.
OUTPUT reporting-narrative — written findings and presentation structure, tuned to the audience and format.
03 — RULES THE VAULT ENFORCES
04 — A REQUEST, FLOWING THROUGH

Take “why did signups drop in June?” The orchestrator refuses to jump to SQL. It runs the spine:

FRAME: METRIC, WINDOW, HYPOTHESES PROFILE THE DATA SEGMENT THE DROP TEST: REAL OR NOISE? REPORT WITH CONFIDENCE LEVELS

Each arrow is a gate: the statistics skill won’t bless a segment difference without a test; the reporting skill won’t write “signups dropped because” when the honest sentence is “the drop concentrates in paid channels — cause unconfirmed.” The proof it works: this entire site — design, copy, generated art, CV pipeline — was produced inside the Vault.

05 — SYSTEM FAMILY & DIRECTION

The Vault shares an architecture family with JobSeek (job-search intelligence) and LifeOS (personal tracking): markdown-first storage, file-based knowledge that accumulates, generated dashboards, and a strict separation between what humans edit and what machines write.

NOW DAILY USE ON REAL ANALYSES — EVERY PROJECT FEEDS SKILLS/ WITH NEW PLAYBOOKS
NEXT EXPERIMENT-DESIGN DEEPENING — A/B PLAYBOOKS, POWER CHECKS, READOUT TEMPLATES
REVIEW PUBLIC RELEASE — METHODOLOGY ONLY, AFTER A FULL PRIVACY PASS

DIRECTION, NOT PROMISES — DATED WHEN SHIPPED.

Want the war stories behind it?

ONE TECHNICAL NOTE IS UP — MORE AS THEY’RE WRITTEN

Read: the ROW_NUMBER hunt ← All systems