Mike Hammett

How things actually work.

The Pipeline, and Why This Work Comes Last

Written by

in

More is coming. Fourteen pieces are sitting in the queue right now, plus the eleven I’ve already gotten out the door. Some have real draft text already. Others are still research and structure, not yet prose. A few are older drafts that need a fact-and-currency check before they’re usable, not just an edit pass.

None of it is close to my top priority. I have a day job. I have a side job. I have a family. It gets whatever time is left after everything else that actually has a deadline.

This site is a real expense. No ads. No sponsors. I could run either. I won’t. I want unadulterated facts, and the fastest way to lose that is to owe somebody something. I name my employer and my side job, and disclose how they relate to this coverage, on the About page.

It’s not writer’s block. It’s queue position.

I’m building a verification pipeline, and it isn’t finished yet. That’s why things take longer. Once it’s running end to end, it will check every claim in every draft before anything goes near WordPress. It will check citations against primary sources, not against whether they sound plausible. It will stress-test arguments to find the weak points I didn’t catch. It will flag drafts wherever I stated something with more certainty than the evidence supports.

I’m still building and testing it. It’s not a plugin I installed. It’s a multi-stage review process I designed and directed Claude to build, and it runs each draft past models from six different providers, not Claude alone. A verdict only counts if it comes back with a verbatim quote and a source URL. On the articles here, Claude drafts. I set the voice, correct the facts, and decide what survives. A model drafting doesn’t make a draft verified, and neither does a person drafting. I check the facts against sources, regardless of who wrote the sentence. Before the checker existed, that was all manual: copying and pasting drafts between AI platforms. Some of the eleven likely went through earlier versions of the checker. The code is already public on GitHub, but it isn’t done. Once it’s solid, I’ll say so, and anyone else running a fact-based publication can use it too.

The fast version of this job is prompting an LLM and hitting publish. I’m running the slower version: draft, then verify. That’s on purpose.