convez.ai
AgenteDados & AnalyticsAvançado
link7373

agentic-bi-team

link7373/agentic-bi-team

Time BI virtual pronto para usar: orchestrator + 8 especialistas + workflows em Markdown

Ver no GitHub
InicialPluginScore · de 0 a 100

A complete Agentic BI & data analytics team for Claude Code. Fill out one charter, run /setup-team, and you have 8 specialists, 10 workflows, and a full knowledge base.

Instalação

$npx convez-ai link7373/agentic-bi-team

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
4
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 1 semana
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Orquestrar pipelines de dados (ETL) sem infraestrutura externa
  • Gerar análises ad-hoc e scorecards de performance automaticamente
  • Criar dashboards e relatórios executivos em múltiplos formatos
  • Implementar anomaly detection e análise de causa-raiz em KPIs
  • Construir modelos de ML (churn, propensity, forecasting) como parte do workflow

Veredito

Solução ambiciosa e bem estruturada para equipes de dados que usam Claude Code. A arquitetura em 3 camadas (orchestração + especialistas + memória) é sólida, e o escopo cobre quase todas as funções de um time BI real. Requer disciplina no preenchimento do charter inicial e leitura atenta da documentação, mas elimina handoffs manuais.

README

A tradução para o português deste README ainda está na fila.

Agentic BI Team for Claude Code

A complete, self-sustained Business Intelligence & Data Analytics team that you run with Claude Code. Fill out one plain-English charter document, run one command, and you have a configured virtual team that covers everything a real BI team does:

  • ETL & data pipelines — ingest from source systems, layered raw→staging→marts models, quality gates
  • Summary tables from huge datasets — fast, pre-aggregated tables at the right grain
  • Ad-hoc analysis & short reports — business questions answered with reproducible queries
  • Cross-database joins — complex analysis spanning CRM + billing + product data, with match-rate honesty
  • Weekly & monthly performance scorecards — fixed KPI set, status colours, narrative explanations
  • Dashboards — Tableau, Power BI, Looker, or self-contained HTML, with consistent design standards
  • Proactive monitoring — anomaly detection and structured root-cause analysis when metrics underperform
  • KPI & metric definition — industry-best-practice metric trees with anti-gaming guardrails
  • Machine learning — churn/propensity models, forecasting, segmentation, A/B test design, model cards
  • Executive deliverables — PowerPoint, Word, Excel (and Google Workspace equivalents) built from analysis
  • Domain learning — researches your industry, benchmarks, and competitors, and remembers what it learns

Everything is plain Markdown. No app, no SaaS, no lock-in — the "team" is a set of instruction files that Claude Code reads.


How it works

                    ┌─────────────────────────────┐
   You (plain      │  CLAUDE.md  (Head of BI)     │   routes work, QAs output,
   English) ─────► │  reads knowledge/, applies   │   delivers finished results
                    │  standards/, delegates to:   │
                    └──────────────┬──────────────┘
        ┌──────────┬──────────┬────┴─────┬──────────┬──────────┐
        ▼          ▼          ▼          ▼          ▼          ▼
   data-      analytics-   bi-      dashboard-  data-     insights-
   engineer   engineer     analyst  developer   scientist communicator
        (+ metrics-steward and performance-monitor)

Three layers make it self-sustaining:

LayerFilesRole
OrchestrationCLAUDE.mdThe Head of BI: business context, routing rules, operating principles, escalation rules. Claude Code loads this automatically every session.
Specialists & workflows.claude/agents/*.md, .claude/skills/*/SKILL.md8 subagents with deep role-specific instructions; 10 slash-command workflows with step-by-step procedures.
Memory & standardsknowledge/*.md, standards/*.mdThe team's long-term memory (data quirks, metric definitions, stakeholder preferences, decisions) and house style. Agents read these before working and write back what they learn.

Repository layout

├── CLAUDE.md                  Master orchestrator (auto-loaded by Claude Code)
├── START-HERE.md         ← YOU fill this out, once, in plain English
├── README.md                  This file
├── .claude/
│   ├── agents/                8 team members (Claude Code subagents)
│   │   ├── data-engineer.md          ETL, ingestion, data quality
│   │   ├── analytics-engineer.md     summary tables, marts, metric layer
│   │   ├── bi-analyst.md             analysis, cross-DB joins, deep dives
│   │   ├── dashboard-developer.md    Tableau / Power BI / Looker
│   │   ├── data-scientist.md         ML, forecasting, experiments
│   │   ├── metrics-steward.md        KPI definitions & governance
│   │   ├── performance-monitor.md    scorecards, anomalies, root cause
│   │   └── insights-communicator.md  decks, docs, spreadsheets
│   └── skills/                10 workflows (slash commands)
│       ├── setup-team/        /setup-team — configures everything from your charter
│       ├── build-pipeline/    /build-pipeline — ETL & summary tables
│       ├── analyze/           /analyze — business-problem analysis
│       ├── scorecard/         /scorecard weekly|monthly
│       ├── build-dashboard/   /build-dashboard
│       ├── investigate-metric//investigate-metric — anomaly & root cause
│       ├── define-kpis/       /define-kpis
│       ├── build-model/       /build-model — ML development
│       ├── make-deliverable/  /make-deliverable — slides / docs / Excel
│       └── research-domain/   /research-domain — industry & product learning
├── knowledge/                 Long-term memory (populated by /setup-team, grown over time)
│   ├── business-context.md    company, priorities, glossary, change ledger
│   ├── data-sources.md        connections, table inventory, join keys, landmines
│   ├── metrics-catalog.md     THE source of truth for every metric definition
│   ├── stakeholders.md        audiences, preferences, distribution rules
│   ├── industry-notes.md      research briefings & benchmarks
│   └── decision-log.md        methodological decisions & observations
├── standards/                 House style applied to all output
│   ├── sql-and-data-standards.md
│   ├── dashboard-standards.md
│   └── reporting-standards.md
└── analytics.md               Standing analytical framework — statistical reasoning & visualization principles

Working artifacts are created as the team operates: analyses/, pipelines/, dashboards/, models/, scorecards/, deliverables/.


Quick start

Prerequisites

  • Claude Code (CLI, desktop, or web)
  • Some way for Claude Code to reach your data — a CLI client (psql, bq, snowsql, sqlite3...), an MCP server for your warehouse, or even just CSV files in the repo. Don't worry if this isn't set up; /setup-team will help you sort it out.
  • Optional: Python 3 (used for ML and for generating .pptx/.docx/.xlsx deliverables — packages are installed on demand)

1. Get the kit

git clone https://github.com/link7373/virtual-bi-team.git my-bi-team
cd my-bi-team
claude

The .claude/ folder must be at the root of the workspace Claude Code opens — that's how the agents and skills load. (Run /agents inside Claude Code to confirm the 8 team members are visible.)

2. Fill out the charter

Open START-HERE.md and answer its questions in plain English — bullet points and brain dumps are fine, no technical vocabulary needed. It covers seven areas:

  1. The business — what you do, how you make money, top priorities, the one number that matters
  2. Your data — where it lives, roughly how big, how to connect, known problems
  3. Metrics & reporting — what you measure today, what you wish you could answer, who reads reports, scorecard cadence
  4. Tools & outputs — dashboard tool (Tableau / Power BI / Looker / none), deliverable formats, branding
  5. Advanced analytics — predictions you'd value, Python environment
  6. Rules & boundaries — privacy/compliance rules, things the team must never do without asking, things it's pre-authorised to do
  7. Context & quirks — fiscal year, timezone, company jargon, history a new hire should know

Leave blank anything you don't know. The setup process asks about gaps and applies sensible defaults for the rest.

3. Run setup

In Claude Code:

/setup-team

This reads your charter and:

  • asks you one batched round of clarifying questions for anything ambiguous,
  • tests every data connection you named and records exactly what works (and what's blocked, and why),
  • discovers your database schemas and builds a table inventory,
  • replaces every {{placeholder}} across all files with your real configuration,
  • seeds a draft KPI list appropriate to your business model, and
  • reports back with a suggested first task.

From this point the team is live.


Configuration reference

The placeholder system

Every file ships with {{PLACEHOLDERS}} marking where your configuration goes — {{COMPANY_NAME}}, {{BI_TOOL}}, {{DATA_PRIVACY_RULES}}, scorecard thresholds, brand colours, and so on. /setup-team fills them from your charter. To find anything still unconfigured:

grep -rn "{{" --include="*.md" .

You can also edit any file by hand at any time — they're just Markdown, and hand edits are equally authoritative.

Connecting your data

The team reads connection details from knowledge/data-sources.md. Three patterns, in order of preference:

  1. MCP server for your warehouse (BigQuery, Snowflake, Postgres, etc.) configured in Claude Code — gives the team direct, permissioned query access.
  2. CLI client available in the shell (psql, bq, snowsql, duckdb, sqlite3) with credentials in environment variables — the team records the exact working command.
  3. Files — CSV/Parquet exports dropped into the repo; the team will analyse them with DuckDB/pandas and tell you what regular exports it needs.

/setup-team verifies whichever you have with live test queries; nothing is recorded as "working" untested.

Configuring the BI tool

Set your tool in the charter (or directly in CLAUDE.md and dashboard-developer.md). The dashboard developer carries tool-specific build rules for Tableau (published data sources, extracts), Power BI (star schema, thin DAX), and Looker (LookML kept in this repo). If Claude Code has no direct API access to your tool, the team produces complete import-ready artifacts plus exact setup instructions. With no BI tool at all, it builds self-contained HTML dashboards regenerated by script.

Tuning behaviour

Want to changeEdit
Routing, operating principles, escalation rulesCLAUDE.md
What a specific role does or how it works.claude/agents/<role>.md
The steps of a workflow.claude/skills/<name>/SKILL.md
Metric definitions, targets, scorecard thresholdsknowledge/metrics-catalog.md (or run /define-kpis)
SQL conventions, naming, quality gatesstandards/sql-and-data-standards.md
Chart/colour/layout rulesstandards/dashboard-standards.md
Report structure, tone, brandingstandards/reporting-standards.md
Who gets what, in which formatknowledge/stakeholders.md

Safety rails

Two placeholders in CLAUDE.md §8 govern autonomy, set from your charter:

  • Never without asking: destructive actions and outward-facing distribution always require confirmation unless you pre-authorise them.
  • Pre-authorised: things the team may do freely (e.g. "read any table", "create tables in the dev schema").

Privacy rules ({{DATA_PRIVACY_RULES}}) propagate to querying, dashboards, and deliverables, including minimum aggregation sizes.


Day-to-day usage

Just talk to it — the orchestrator routes to the right specialists:

"Why did signups drop last week?" → root-cause investigation "I need a board deck on Q2 performance" → analysis → deck, with an outline shown to you first "Our queries on the events table take forever" → summary-table design and build "Which customers are likely to cancel?" → screening question first (is a model even needed?), then ML workflow

Or invoke workflows directly:

CommandWhat happens
/analyze why is enterprise churn risingFramed, profiled, pressure-tested analysis with a written FINDINGS.md
/scorecard weeklyThis week's scorecard in scorecards/, every red/yellow explained
/investigate-metric signups down 20%Data-problem check → localisation → timeline → tested hypotheses → verdict with confidence
/build-pipeline daily revenue summary from billingDesign doc → layered build → quality gates → docs
/build-dashboard exec revenue overviewSpec (confirmed with you) → data layer → build → number-by-number validation
/define-kpisMetric tree + rigorous catalog definitions + targets, ratified with you
/build-model churn predictionLeakage-safe dataset → baseline → model → business-terms evaluation → model card
/make-deliverable churn analysis as exec deckPyramid-structured .pptx with every figure source-mapped
/research-domain SaaS churn benchmarksSourced, dated briefing in knowledge/industry-notes.md

Running on a schedule

The scorecards and monitoring are designed for recurring runs. Options:

  • Headless CLI: claude -p "/scorecard weekly" from cron or any scheduler.
  • Claude Code web: start a session on this repo and ask for the scorecard; or use a GitHub Action that invokes Claude Code on a schedule.
  • Manual: just run /scorecard weekly each Monday — it takes one command either way.

How the team stays smart

The knowledge files are working memory, not documentation theatre — agents are instructed to read them before tasks and write back after: data quirks land in data-sources.md, definition rulings in metrics-catalog.md + decision-log.md, stakeholder reactions in stakeholders.md, business events in the change ledger (which powers root-cause timelines). Commit the repo regularly; the git history is the team's institutional memory.

Re-run /setup-team after major changes (new warehouse, new product line, reorg) — it merges new facts without wiping accumulated knowledge.


Extending the team

  • Add a team member: create .claude/agents/<name>.md with frontmatter (name, description) + role instructions, and add a row to the routing table in CLAUDE.md §3. Useful additions: a financial analyst, an experimentation specialist, a data-governance officer.
  • Add a workflow: create .claude/skills/<name>/SKILL.md (frontmatter + numbered procedure) and list it in CLAUDE.md §4.
  • Industry packs: the metrics-steward and /define-kpis adapt to your business model from the charter; for deep vertical needs, extend knowledge/metrics-catalog.md and industry-notes.md directly.

Troubleshooting

SymptomFix
Agents/skills don't appear.claude/ must be at the root of the folder Claude Code opened. Check with /agents.
Team asks for context it should knowPlaceholders left unfilled — run the grep above, or re-run /setup-team.
Two reports disagree on a numberThat's a metrics-steward job: say "these two numbers disagree" and it will reproduce both, rule, and fix the deviating artifact.
Data connection brokeUpdate knowledge/data-sources.md (or tell the team — it will retest and update the file).
Output style isn't rightEdit the relevant standards/ file once; every future artifact follows it. Log stakeholder-specific reactions in knowledge/stakeholders.md.

License & disclaimer

Use freely. The kit is instructions, not advice — validate business-critical numbers and decisions the same way you would coming from any analyst.

Ver o original no GitHub

Bom para

Relacionados