ContractorKeith “delve” used to be a perfectly good word. then the language models got hold of it, and now you can’t use it without sounding like a press release. same for “tapestry.” same for “landscape” when nobody is talking about actual land.
i write this blog with help from claude. that’s not a secret, it’s the premise of the whole site. but i want the posts to sound like me, a guy who spent 37 years on jobsites, not like the average of every linkedin post ever written. so in march i built a claude code skill that forces the issue.
the tells are documented
wikipedia editors maintain a page called “signs of AI writing.” it’s roughly 1,600 lines of patterns collected by people who clean up machine-generated articles all day. researchers have published on the same thing: kobak et al. (2025) tracked how words like “delve” spiked in academic writing after 2022, in papers that real humans put their names on.
my favorite find was the guardian’s name for synonym-cycling: the “popular orange vegetable” problem. an AI writing about carrots will call them “the vegetable,” “the produce,” “the root crop,” anything but carrots, because the model gets penalized for repeating itself. humans just say carrots.
and the plainest tell surprised me most. one study measured “is” and “are” dropping more than 10% in academic writing after 2022, because models keep swapping them for “serves as” and “stands as” and “represents.” turns out the most boring words in english are the most human ones.
44 banned words and 6 hard checks
the ai-writing-skill encodes all of this as negative constraints. 44 banned terms sorted into three tiers, from instant giveaways (delve, tapestry, pivotal, meticulous) down to context-dependent ones (robust, comprehensive, navigate). structural rules too: no “challenges and future outlook” endings, no opening a post by defining its own topic, no sentences trailing off with “…highlighting the importance of” anything.
then the enforcement side: 6 binary checks every draft has to pass before it ships. at least one sentence fragment. at least two sentences starting with “and,” “but,” “so,” or “because.” real names and numbers instead of “experts say.” pass or fail, no partial credit.
testing it like code
a skill file is just instructions, and instructions drift. you don’t find out whether they hold by reading them. you find out by running them.
i borrowed the approach from karpathy’s autoresearch repo: define binary eval criteria, generate articles against 8 different test prompts (tutorials, opinion pieces, product comparisons, a personal essay), score every output, tweak the skill, run it again. the target is 27 of 30 checks passing, held across 3 consecutive runs. a full optimization cycle costs somewhere between $6 and $12 in api calls, which is less than an hour of any editor i’ve ever met.
that loop is the part worth stealing if you build skills of your own. the skill file is the easy half. the eval harness is what makes it trustworthy.
does it work?
you tell me. every post on this site goes through it, including this one. if the writing here sounds like a person, the skill is doing its job. if it doesn’t, the eval suite and i have more work to do.
the repo is public. drop it in ~/.claude/skills/ and your claude picks it up automatically. fair warning, though: once you learn the tells, you see them everywhere. linkedin becomes unreadable.
you’ve been warned.