A machine drafts. I decide.
Every piece on this site passes through the same five steps. Here they are, including the one that is mine alone, and the months when the whole thing stopped.
I built this pipeline in April 2026 to run the marketing of a small business, then pointed it at my own writing to see whether it could hold a standard. The honest answer is in the pieces, and in the gap between May and October 2026 when it published nothing. The steps below are the version that runs now.
A job drafts
On a Monday morning a scheduled job on my own machine picks the pillar that has waited longest, researches the week with live sources, and writes a draft of 1,200 to 1,800 words. On Thursdays it writes a shorter signal note about one thing that happened.
It writes against a set of standing rules I keep in a plain text file: what the site is, who it is for, what is banned, how a piece is built. The same file carries every rule the pipeline has learned from my edits.
A critic checks
A second pass scores the draft against those rules and against one question: would a working professional send this to a peer who has never heard of me? It also checks that every figure links to a source in the same paragraph, and that every link resolves. A draft that fails is rewritten once, then stops.
A pull request waits
The draft lands as a proposed change to this site, with a preview of the finished page, the critic's report and the list of sources. Nothing is published. It sits there until I read it.
I edit, or I decline
This is the step that is only mine. I read the preview on my phone, change what needs changing in the text itself, and either merge it or close it. Merging is publishing: the site rebuilds from the merged text and goes live within a minute.
The pipeline learns from the difference
Within the hour a job compares what was proposed with what I merged. Each change becomes a dated note; a change that keeps recurring becomes a proposed rule for me to accept or delete. Sources I removed go on a list the drafts avoid. That file, not a model, is the memory of this site.
Automated
- Research with live sources, every run
- Drafting of articles, notes and the monthly email
- Source checks: a link for every figure, every link resolving
- Rendering of every page, listing and sitemap from one manifest
- Learning from the difference between the draft and the merge
Mine
- The argument a piece is allowed to make
- Every edit, in the text, before it goes live
- The decision to publish, or not
- The rules the pipeline writes against
- Sending the monthly email, by hand
The stall
The first version of this pipeline needed me to approve each stage inside a web interface I had also built. The scheduled job on my laptop never ran once: the operating system refused it access to the folder. A copy on a server kept drafting every Monday into a queue nobody opened. From May to October 2026 the site published nothing, while the drafts piled up behind a gate I had stopped visiting.
The fix was not a better interface. It was removing the interface: the draft now arrives where I already read things, as a proposed change with a preview, and my edit is the signal. If that sounds like a lesson about automation in general, it is the one this site exists to keep learning in public.
The hero images are the one flourish. Sur Bacchus, a tabby with high standards, sits at the top of each article in a scene drawn from the piece. He does not write anything.