Claude Code

5 posts in Blog

Claude Code is what I reach for when a job is too big to do by hand and too specific to buy software for.

These are the posts where it did real work: a six-book series written and voiced end to end, a month of usage broken down to find where the tokens actually went, and a plain account of how Claude Code and Cowork split the work on a normal day.

I have tried to leave the failures in. The interesting part is rarely the finished output, it is the three attempts before it.

Three Claude accounts, one Mac, and the one thing I couldn't fix

Published on Sep 14, 2026

Claude Code authenticates one account per machine and the desktop app holds one login at a time. Both key their credentials to a directory, which turns out to be the whole trick. Here is the setup for a separate account per client, the traps, and the part that is genuinely impossible.

My agent loop was 62% of a day's tokens. I blamed the code reviews.

Published on Sep 5, 2026

An automated agent loop ran 30 iterations inside one conversation and used 62% of a day's tokens. The code reviews I was suspicious of were half a percent. Here is the mechanism, and what actually fixes it.

I measured where my AI budget actually goes

Published on Jul 31, 2026

I had a confident theory about which projects were expensive. I was wrong four separate ways, and the thing I was rationing turned out to be one percent of the total.

The six-book series I've never read

Published on Jul 31, 2026

I used Claude Cowork to write a six-book LitRPG series and render it as a cast audiobook. Four hours of finished audio cost $22.09.

How I actually use Claude Code and Cowork

Published on Jul 14, 2026

Where AI agents actually fit in my work: content pipelines, batch jobs, and API testing, plus the parts I still do myself.