Chief of Staff — Assistente Agentico de Produtividade
bcgwynn/chief-of-staff-build
Assistente que centraliza tarefas e surfata ações prioritárias sem você precisar procurar.
Ver no GitHubUm agente autônomo que centraliza tarefas, notas e ideias espalhadas em múltiplas ferramentas (Apple Notes, Google Docs, Granola) e as organiza de forma estruturada. Funciona integrado ao Slack como um sistema de "push" — ele vem até você com resumos diários, extrai ações de reuniões automaticamente e gerencia prioridades via Google Sheets.
Instalação
npx convez-ai bcgwynn/chief-of-staff-build1 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
Casos de uso
- Consolidar notas e tarefas dispersas em um único hub
- Receber resumos diários de ações prioritárias no Slack
- Extrair automaticamente action items de transcrições de reuniões (Granola)
- Manter rastreamento de tarefas e ideias de conteúdo sem friction
Veredito

Implementação madura e prática de um agente que resolve um problema real — fragmentação de contexto. A arquitetura é clara, bem documentada e usa bem as capacidades do Claude Code. Ideal para quem tem fluxo de informação disperso e quer um sistema que proativamente organize e priorize.
README
A tradução para o português deste README ainda está na fila.
Chief of Staff — Build Log
An agentic chief of staff built with Claude Code, Google Sheets, Slack, and Granola.
Why I built this
I had scattered notes everywhere — Apple Notes, Granola, Google Docs — with no reliable system for surfacing action items or follow-ups. Things got captured but not completed. LinkedIn content ideas were buried in the same pile. I was doing all the context-holding myself.
The chief of staff is my attempt to offload that. It runs on a schedule, proactively surfaces what matters, and gets out of the way. The core design principle: push, not pull — the system comes to me, I don't go looking for it.
What it does
- Daily nudge — every morning, scans my action item tracker and brain dump channel, then posts a structured summary to Slack with priority tables, anomaly flags, and a 2-3 sentence analysis
- Brain dump processing — anything I post to a dedicated Slack channel gets routed: tasks → action items, ideas → content ideas, notes → saved for later
- Meeting processing — post
process [meeting name]in Slack and it pulls the Granola transcript, extracts action items, flags content-worthy moments, and posts a summary - Command processing — post
done AI-007ormark AI-003 in progressin Slack and it updates the tracker automatically
Architecture

| Tool | Role |
|---|---|
| Claude Code | The brain — reads, reasons, writes |
| Google Sheets | Persistent action item store (Action Items tab + Content Ideas tab) |
| Granola | Meeting notes source |
| Slack | Delivery channel (nudges, summaries) + input (brain dump, commands) |
| Google Apps Script | Webhook for all Sheet writes |
| launchd | Automated scheduling on macOS |
Data flow:
- In: Granola (meetings) + Slack #brain-dump (everything else)
- Store: Google Sheet
- Out: Slack #cos-updates (nudges, summaries) + Google Calendar (deadline events)
What's built (v1)
- Daily nudge with structured format — per-goal tables, 🔴🟡🟢 priority, ⚠️ anomaly flags, analysis summary
- Brain dump routing with content type detection (action item vs. content idea vs. observation)
- Meeting processing via Slack command (
process [meeting name]) - Slack command processor — handles status updates, new items, content idea updates
- Apps Script webhook — append rows, update AI- items, update CI- items, add content ideas
- Bot identity — posts as Chief of Staff APP, not my personal Slack profile
- Automated via launchd — nudge fires on two triggers: login-triggered LaunchAgent (fires on restart) and sleepwatcher LaunchDaemon (fires on wake from sleep). A once-per-day guard is shared between both triggers — whichever fires first wins, the other skips. Command processor runs at noon / 4pm via StartCalendarInterval. If Mac is asleep at those times, that run is skipped; items are processed on next trigger. Nothing is lost permanently.
What's next (v2)
- Meeting processing automation (no manual trigger)
- Meeting filter by Granola folder
- Evals framework — scoring nudge quality, extraction accuracy, content idea hit rate
- Weekly content scan + progress snapshot
- Meeting processing confirmation step — AIs require approval, CIs auto-written as Draft
Files in this repo
CLAUDE.md— the system prompt / operating instructions for the CoS. This is what Claude reads at the start of every session to understand goals, tools, and how to behave.daily-nudge.sh— shell script triggered on login (RunAtLoad). Runs the command processor first, then reads the Sheet and posts the nudge. Once-per-day guard prevents re-firing on re-login.process-commands.sh— shell script run by launchd at noon and 4pm (StartCalendarInterval). Also called synchronously by the nudge at login so brain-dump is processed before the Sheet is read.chief-of-staff-brief.md— full project brief: architecture decisions, roadmap, and lessons learned.
Build log
Session-by-session decisions, tradeoffs, and failures are maintained separately outside this repo. The failures are documented intentionally — this is a build-in-public project and the mistakes are part of the story.