Problem
I use Claude for research, session logs, briefs, and first passes on fiction. Left to itself, a model writes in a register anyone can spot: "This isn't X, it's Y," three adjectives in a row, "it's worth noting," a participle phrase tacked on to fake depth. A style guide written for people ("keep it warm, avoid clichés") does nothing for a model, because it can't be checked. I needed rules a program can apply and fail.
Role
Author. The rules came from my own reading of what gives machine text away, from peer-reviewed detection work and Wikipedia's AI-writing field guide, and from the tells I kept catching in drafts. Claude drafted wording from my notes; I kept, cut, or rewrote every line and set every ruling.
What I built
A single skill file, 2,616 words, that loads at the start of every session in every project. It has three parts.
Part 1 is how the model talks to me in chat: contractions, direct address, a stance, no padding. Part 2 is the substance rules for anything produced under my name: 26 numbered rules in four groups, each with the pattern, the disguises it wears, and the fix. Part 3 is an audit mode that runs only when I ask: every hit cited with the rule, the line, the span verbatim, and a suggested fix, and nothing edited until I rule on it.
The rules are written so a hit is unambiguous. The first and most important one:
1A. Negative parallelisms and reframe constructions (FATAL)
The single most reliable tell of AI-generated text. If you see even ONE in your output, rewrite the entire sentence.
The pattern, in all its disguises: "This isn't X. This is Y." "Not X. Y." "Less X, more Y." "Not only X, but also Y." "It's not about X. It's about Y." "The question isn't X. The question is Y." "Nobody talks about X." "Despite its [positive], [subject] faces challenges..." Any sentence that negates a framing then asserts a corrected one.
The fix: delete everything before the positive claim. "It's not about the prompt. It's about the context" becomes "It's about the context." The negated framing adds zero information.
One rule, applied to the kind of sentence a model writes for a session log, before and after:
Before. This session wasn't about fixing the build. It was about finding out why the build kept breaking in the first place.
After. This session found out why the build kept breaking.
Every rule carries a strength: banned words and structures at 80 to 90 percent, sentence variety and direct address at 60 to 70, word choice left to context. The guide says in its own text that a banned word is allowed when it's right for the subject, so the model has room to judge.
Limits
The guide catches patterns. A clean paragraph can still be dull. It also drifts: the model learns to avoid the listed tells and finds new ones, so the list gets revised whenever I catch a new tell by reading; the latest ruling, American English throughout, landed 10/13/2026. And the audit is only as good as my ruling on each finding; the skill is built so that nothing is edited without one.
Result
Every piece of text that leaves my projects passes an inline check against Part 2 before delivery, and I can ask for the full audit on any file with one word. The guide is the source for 6 other skills (coursework, manuscript, screenplay, and the rest), which specialize it rather than restate it. The same file is what this portfolio was built against.