From Operator to Architect: Building an AI Chief of Staff with Claude Cowork
Every solo founder hits the same wall: you are simultaneously the CEO, the engineer, the support team, and the operations person. The tasks that eat the most time are not the hard ones — they are the repetitive, low-judgment ones. Reading the same categories of email. Triaging the same types of issues. Checking the same dashboards. Claude Cowork is my attempt to delegate that layer.
The 5-Layer Architecture
Cowork is not a chatbot. It is a structured workflow system built across five layers:
- Data — Raw inputs: Gmail threads, calendar events, Supabase query results, Notion pages
- Processing — Normalization and enrichment: parsing emails into structured objects, tagging urgency
- Skill — Domain-specific logic packaged as reusable CLAUDE.md skill files
- Orchestration — The layer that chains skills together and decides what runs next
- Interface — How results surface: Slack digest, email draft, terminal output, or Notion update
Each layer has a single responsibility. Adding a new skill means writing one file — the orchestration layer handles the rest.
The Inbox Command Center Skill
The first skill I shipped handles email triage. It reads a batch of Gmail threads, classifies each by urgency and category, drafts responses for anything that needs a reply, and writes a prioritized digest.
The Gmail API integration pulls threads via OAuth. Claude reads each thread in full context — not just the subject line — and applies a classification rubric:
- P0: Needs same-day action (customer issue, payment problem, legal)
- P1: Needs response within 48 hours (partnership, hiring, vendor)
- P2: FYI or async (newsletters, receipts, digests)
Urgency classification runs locally. No email content leaves for a third-party classifier — it goes directly to the Claude API under my key.
What This Unlocks
Before Cowork, my morning started with 30 minutes of inbox triage. That time now goes to building. The pattern is transferable: any repeated judgment task — sprint planning, metrics review, customer feedback synthesis — can become a skill with a well-defined input schema and a clear output contract.
The architecture is intentionally boring. Skills are markdown files. Orchestration is a TypeScript runner. The power is in the compounding: each new skill makes the whole system more useful without touching the layers below it.