I Gave an AI Agent My To-Do List for a Week

I run a publishing company, write books, maintain a website, and hold down a full-time cybersecurity career. My to-do list isn't aspirational. It's a survival mechanism.

In August I decided to run an experiment: hand as much of that list as possible to Claude through Cowork, the desktop agent that Anthropic has been building out. Not as a gimmick. As an honest test of whether an AI agent can carry real operational weight for a solo operator.

Here's what worked. Anything with a clear input and a defined output. "Take this blog post content and publish it to the website." "Generate a book metadata description from this manuscript chapter." "Research KDP formatting requirements for hardcover trim sizes." These tasks went smoothly because they have boundaries. The AI knows when it's done because the definition of done is explicit.

Here's what didn't. "Review the manuscript for consistency." That sounds defined, but it isn't — consistency against what standard? What counts as an inconsistency versus an intentional choice? The AI flagged dozens of things. Some were real catches. Many were stylistic preferences it was projecting onto my writing. I spent more time reviewing its reviews than I would have spent just doing the review myself.

The lesson isn't that agents don't work. It's that they work exactly as well as your instructions. When I gave Cowork a tight scope with clear success criteria, it was genuinely useful — it saved me hours on tasks I'd been putting off because they were tedious. When I gave it something fuzzy and said "figure it out," I got fuzzy results that needed heavy editing.

The biggest surprise was the compound effect. By the end of the week, I'd built up a set of task patterns that I could reuse. The agent wasn't learning from week to week, but I was getting better at defining tasks in a way that produced consistent results. My prompts got shorter and more precise. My output expectations got more realistic.

Would I do it again? I am doing it again. Right now. The experiment became the workflow.