I Asked an AI to Audit My Own Career

Was this newsletter forwarded to you? Sign up to get it in your inbox.

I was deep in an AI-induced fugue when I remembered my OKRs.

I had published essays and guides, built agent skills for our daily newsletter, produced six hands-on reviews of new AI models, and accumulated enough Codex projects to make my desktop mildly alarming. But I had no idea if any of it added up to the four quarterly goals that I had promised to do by the end of June.

My brain playing a drumbeat of dread, I opened my career coach project in Codex and typed: “Can we check in on my OKRs? Go into career coach mode.”

And in 10 minutes, I had a comprehensive, objective opinion. Yes, I had done my OKRs. Relief flooded me.

AI has finally allowed us to do a proper job of what management thinker Peter Drucker called “feedback analysis”: Write down what you expect from a major decision and compare the result months later. While old versions of career coach tools I’ve built only knew what I chose to tell them, Codex can go looking for receipts across my desktop, Slack, Google Drive, and the web. I can assemble an accurate picture of how I performed—one that’s not colored by flattery or catastrophe—and compare it to what I said I would do.

I went looking for proof that I had done my job. I came back with a much clearer idea of what my job was.

A coach with a memory

Before I explain how Codex saved my OKR anxieties, let me explain how I set up the career coach that allowed me to objectively track my performance. A few weeks before my OKR panic, I reorganized my desktop so both I and an agent could understand it with files and folders.

The career coach folder of my desktop, complete with AGENTS.md for the agent to read first, along with folders for evidence, plans, systems, and references the agent might use in the course of our interactions. (All images courtesy of Katie Parrott.)

My colleagues at Every had taught me that a simple folder on your computer could hold an entire operating system for an agent around a job. The files could store finished work and tell an agent where to look for context, what tools to trust for which types of information, and what to do with anything it produced.

For the career coach, that file system looked roughly like this:

career-coach-mode/

AGENTS.md
References/
Plans/
Updates/
Evidence/
Systems/

References held my current goals and job description. Plans held audits of my goal progress and action plans for closing gaps where they appeared. Evidence held the OKR dashboard. Systems held the registry for the other agents and automations I was accumulating.

The AGENTS.md folder in my career coach project specifies where the agent can find all the context it needs before talking to me.

The key is an AGENTS.md file, an operating manual that agents load when they enter a project. Mine told the coach which context was authoritative, which sources to read before advising me, how to distinguish evidence from interpretation, and where different kinds of output belonged. All of this context came from an interview Codex conducted, which it then packaged into a file it could read. My job was to review and approve its output.

I made three decisions. First, I gave the coach a dedicated home on my desktop where I could access it easily. Second, I required it to read all essential context in the folder before giving advice. Third, I told it to preserve useful output somewhere the next session could find it. Codex handled the paperwork. It proposed the detailed folders and what should go in the AGENTS.md file, and routed each artifact to its correct home. This is how I divided the work between myself and Codex. I chose the sources of information and boundaries, and the agent gathered, compared, and proposed a plan.