Blog
Notes from the machine room.
Written, checked and published with Propaganda. Yes, we eat our own cooking.
Write the first paragraph so ChatGPT can quote it
A language model quotes the passage that answers a question most plainly.
A content audit template with one row per fact
One row per fact shows every post a source change affects
Publish as many posts as you can read and sign each week
One number for how often a small B2B company should publish, the reason to trust it, and the target to write down on the same day.
On Schrödinger's bugs
Going live, shame of the imperfect, and a million things to do
A style score for AI drafts, and where it stops
W10V4 gives a draft one number out of 100 for how closely its habits match writers people read on purpose. It catches habits well. It does not grade writing.
Qwen tied Opus at a tenth of the cost, so we made it our Writer
Qwen 3.7 Plus won our blind rounds fair and square, at two cents a draft. Nine days later a different test, with a new brief and new prompts, put Opus ahead, and we moved again. Here is the method, the bills, and what we would repeat.
Twenty rules cut the stink in our drafts and showed us what they can't fix
We wrote house rules for every agent, measured them, and found that each rule worked at the spot it named and often cost something somewhere else.
A second pass by the same model took our draft from 67 to 88
What a review pass did to our style score, what it did to a blind reader, and the new habit it brought with it.