Physician, heal thyself, except the physician wrote the textbook, handed it to the intern, and the intern still botched the diagnosis. I built JAMESISMS.md two weeks ago specifically to stop the author-womack-post skill from sounding like a SaaS landing page with my byline on it. Then I read the twenty posts it produced against the doc it was supposedly reading, and folks, they don’t sound like me. Almost none of the actual vocabulary showed up, no w/, no folks (the irony is not lost on me), no et al, no “I digress,” no clichés landed straight-faced, no sign-offs, nothing from the corpus’s actual word list. Just competent, punchy, entirely generic tech-blog prose with the AI tells filed off.

Went back and read SKILL.md next to JAMESISMS.md to see why, and the answer was sitting right there once I looked for it: the skill had a whole section of banned phrases and banned punctuation, checked by a real script, quality-gate.sh, every single time. It had nothing that required using anything from the doc it was supposed to be voiced by. I’d built a filter and called it a voice. Avoiding em dashes gets you to “not obviously AI.” It does not get you to “sounds like James,” and I’m an idiot who apparently needed twenty published posts to notice the gap between those two bars is the entire point of the exercise.

1
2
3
4
5
6
7
8
9
10

$ jamesisms-check.sh source/_posts/When-Zero-Means-Zero.md   # the pre-fix draft
contractions              8
"w/" for with              0
parentheticals             3
question marks             0
first-person density       0
ADVISORY: zero first-person words (I/my/me) on a 687-word post. This is
personal writing, not documentation, a post can have contractions and
parentheticals and still read like a product changelog if nothing in it is
actually written from a person's point of view.

8 contractions, 3 parentheticals, and still zero I/my/me in almost 700 words. That’s the actual fingerprint of the problem, not the em dashes, the complete absence of a person standing behind the sentences. Fixed SKILL.md so drafting step now requires picking at least 3 specific items from JAMESISMS.md, from at least two different sections, before a single sentence gets written, not after. Added jamesisms-check.sh, which counts contractions, w/, parentheticals against the real corpus’s baseline rates, advisory only, since a script can count whether a word showed up but it can’t tell you if a joke landed. And since JAMESISMS.md’s #references section calls out actual reaction images and GIFs as a real device in the old posts, not just a described one, added verify-image-url.sh, so a meme in a post has to be a real, resolved URL, found for real and checked, never guessed at or remembered wrong. Used it for the first time on this very sweep, that whack-a-mole gif a couple posts back is a real, verified giphy.com link, not a plausible-sounding one I made up.

Then went back through all twenty posts by hand and actually used the vocabulary this time. You’re reading the result of that pass now, in a post about the pass, which is either a little too neat or exactly as neat as it should be, I can’t decide which.