How I work
The operating stack
Plenty of resumes claim "AI-native." This page shows the actual stack. These are the systems that run my own work every day, and every case study on this site was planned, built, or reviewed inside them.
- Second brain / AgentOS
A PARA-organized knowledge base with an agent layer on top. Tasks, projects, and reference material live in one system; agents capture, classify, and route new input so nothing depends on memory or willpower. I designed the same architecture for a commercial coaching product years ago; this is the version I actually live in.
- PRD-first discipline
Every build on this site started as a written PRD before any code. It covers goals, constraints, and acceptance criteria, authored with a packaged PRD skill I run day to day. I check the build against that document as it comes together, the same delivery discipline Microsoft trained into me, just without a team behind it now.
- Skill authoring
Writing voice, PRD authoring, and technical writing all started as things I kept doing by hand, over and over. Each one is now an evidence-based skill spec that an AI can execute and I can review. I write a draft, review the output, refine the spec, and repeat until it matches the standard I'd hold a person to. Think of it as managing a direct report, except the report is software.
- Claude Code as build partner
The voice-capture pipeline, the MCP skills server, and the triage automation all came out of this loop. Each one went through the same cycle. I write the plan, Claude Code batches the execution, I validate the output against the plan, then iterate on what's wrong. Real observability and eval harnesses back every one of them; none of them are demo scripts. I built this site itself the same way.
- Cadence
An agent runs the review mechanics. I keep the judgment call. Daily, weekly, monthly, and quarterly reviews run on a fixed loop. Each level rolls up into the next: daily notes feed the weekly review, weekly feeds monthly, monthly feeds quarterly. The alignment check is built into that same loop. Every task carries a three-bucket score that traces it back to why it matters, so I catch a task's alignment dropping before I've sunk real time into it.
Why this matters to a team
An operator who runs his own work this way can wire the same discipline into an organization. Most "adopt AI" initiatives stall at slideware. I've already done the unglamorous parts: writing the prompts, running the evals, catching the failure modes, and building the operating rhythm around the tools.