We needed one address that could both receive customer mail and send replies from our own domain, without paying for a mailbox provider. The setup is Brevo's free tier, Cloudflare Email Routing on the inbound side, and a Worker in between. Eight DNS records did the job. One of the instructions in the Brevo dashboard would have silently broken every inbound email, and it is the same record every guide tells you to add.

Two AI-drafted novels went live on October 17. One week later the dashboard shows 370 Kindle page reads and no sales — and every one of those pages was read on a single day. The split is the interesting part: 351 pages belong to book one, which means at least two readers finished it, and 19 belong to book two, which means one reader opened it and stopped. Here is what a 24-hour exposure window looks like, and what it does and does not prove.

First rejection on Oct 12: an old personal signup made our company account a duplicate. Fixed that evening at 20:21. Second rejection on Oct 21: thin content, nine days of review. Dates, screenshots, what we change before reapplying, and an update log.

Two full-length novels went through Kindle Direct Publishing last week, drafted by an AI coding agent and packaged by a script. The AI disclosure is not the trap everyone warns you about. The traps are the title collision you only find after you submit, and the table of contents your generator forgot to build.

An always-on AI agent that touches real files and real accounts needs the same review as any junior employee with admin access. We wired ours with three layers: approval gates, hard isolation, and a credential vault the agents cannot read. This is the field guide we wish existed when we started.

Real walkthrough from a mainland company: D-U-N-S came back in Chinese, the US virtual payout account got rejected by SWIFT validation, phone verification was locked until ID approval, and the fix turned out to be a bank trip. October 2026.

Our API-plan quota was burning out weekly on code generation. Splitting the workflow into an agent scheduler plus a subscription IDE cut the drift and moved the cost where it belongs. What we gained in control, what it still costs, and where the split pays off. October 2026.

Publishing a Windows app in 2026 is cheaper and stranger than most guides say: free MSIX hosting, but a trailer uploader that silently fails, a privacy URL that must be live before you even save the Properties page, and two different IDs that look identical. Here is the path we actually took, with the traps marked.

Two home machines need to trade task files and build outputs without a USB stick. Here is the setup that works, with the three things that cost me an evening: the config moved folders, ping lies about the firewall, and discovery is slow on purpose.