How I build: an AI chief of staff with guardrails

Running several production apps alone means a lot of context to carry and a lot of code to review.

• A daily brief. A scheduled agent reads the project list, calendar and inbox every weekday morning and writes a time-blocked plan with the day's three must-dos.
• Shared memory. Every coding session loads that project's notes and decisions first and writes a session note at the end. 580+ notes and 300 session notes across 12 projects.
• 25+ custom Claude Code skills and review agents. How each production app is shipped and reviewed: database-change rules, a security reviewer, a staff-engineer reviewer, SEO publishing, ops lookups.
• Skills for real operations jobs. 18 custom Claude Code skills in the delivery platform's codebase (order lookup, data reconciliation, weekly delivery checklists, database migrations, a ship-to-main checklist), plus 7 reviewer and analyst agents.
• Also: a phone-first rebuild of the company CRM, a fork of CoreLinq's AI CRM (base platform and its voice-calling engine by Matt Rundle). I built the consent engine (STOP blocks everything; a missing marketing opt-in blocks only campaigns), the AI receptionist that answers the business line by voice and text (Case 05), the CRM's MCP server, the phone-first redesign and a database security lockdown, and took the unit tests from 113 failing to 0 (166 test files added).

I design the system and decide what the AI may touch. Claude Code and Codex write most of the code; I write the tests and guardrails, nothing ships without passing CI and my review, and I own what goes to production.

• Memory lives in plain notes the next session reads first, not in anyone's head.
• Reviews are agents with fixed checklists, so every change gets the same scrutiny.

One person running several production apps, with every change reviewed and recorded.