AI Writing Rules

I had a writing-rules document I pasted into project after project. It worked while it stayed in context and stopped working the moment a session got long or I forgot to include it. A document that has to be remembered is a document that gets ignored.

This plugin turns that document into something the harness enforces. Rules load at session start whether or not anyone remembers them, and a checker reads every markdown file Claude writes and reports what it finds. When it finds something, the model fixes it before continuing.

The rules themselves got rebuilt from research rather than from instinct. Wikipedia maintains a page on signs of AI writing that catalogues the tells by era, which matters because the list drifts as models change. "Delve" and "tapestry" dominated 2023. By mid-2025 the common ones were "emphasizing", "enhance", "highlighting" and "showcasing". Alongside that, several groups have measured the patterns rather than only listing them, which is where the thresholds come from.

The finding that changed the design: em dashes are the famous tell and the least useful one now, because everyone scrubs them. Sentence-length uniformity is the tell nobody bothers to fix. Human prose has a burstiness of 0.6 to 1.2, measured as sentence-length standard deviation over the mean. Model output sits at 0.2 to 0.4. That is a number a script can compute, which makes it enforceable in a way that "write with more personality" never was.

Enforcement is the hard half. A skill only fires when the model decides it is relevant, and a model deep in a task does not decide that reliably. The rules had to sit somewhere the model does not get a vote.

Claude Code's hook system solves it, but the exit codes matter more than they look. A PostToolUse hook that exits 0 writes to the transcript and the model never sees it. Exit 2 sends stderr back into the conversation as something to act on. The whole enforcement mechanism is that one difference, so the checker had to be confident enough to justify interrupting a task on every finding.

Confidence meant fighting false positives. The first version flagged code samples, because a shell heredoc full of the word "utilize" is not prose. Fenced blocks and inline code now get blanked before anything is scanned. It also flagged the research files, which is a genuine problem: a catalogue of bad examples is nothing but bad examples, so those files carry an opt-out marker.

Several patterns resisted regex on the first pass. Negative parallelism is the clearest AI sentence shape there is, and "This isn't just a tool, it's a paradigm shift" slipped straight through because the pattern only handled "is not" and never the contraction. Transition stacking failed because "Furthermore" and "Moreover" often open a sentence mid-paragraph rather than a line. The three-item-list threshold of one per 200 words under-fires on a short document, so it scales with word count now.

The last problem was the checker's own credibility. A tool that lints prose has to survive its own linting or nobody will trust it. The first clean run over the research folder found a tricolon I had written two hours earlier, in the file arguing against tricolons.

Three surfaces, each covering what the others cannot.

A SessionStart hook prints 270 words of hard rules into context on every session. No em dashes, straight quotes, sentence case headings, no "not just X, it's Y", break the three-item-list reflex, name the source or drop the claim, plus the vocabulary to avoid and the plain-word substitutions. It costs almost nothing per session and it is there before the first prompt.

A PostToolUse hook runs the checker after every Write and Edit on markdown, mdx, txt and rst. It reports the rule, the line number, the matched text and the fix, then exits 2 so the model has to deal with it. Seventeen regex rules cover em dashes, curly quotes, negative parallelism, the vocabulary list, inflated verbs, dodged copulas, puffed significance, vague attribution, empty openers, wrap-up endings, hedge stacks, filler, assistant scaffolding, leaked model artifacts, Title Case headings, inline-header bullets, and trailing participle commentary. Three statistical rules handle what regex cannot: three-item-list density, transition stacking, and burstiness.

The skill carries the full ruleset and an index into the research, so the model can open the one file it needs instead of loading 29. A /deslop command audits a file, rewrites it, then re-runs the checker to confirm it comes back clean.

The research folder is one file per pattern: what it is, why models produce it, the threshold where one exists, a bad example, the rewrite, and the fix. Keeping them separate means the skill stays small and the detail is there when a specific tell needs explaining.

It runs on Python 3 with nothing from PyPI, and exits quietly if Python is missing so the session-start rules still work. Verified on a sample carrying 14 deliberate tells, where it found all 14 and stayed silent on clean prose. The README passes its own checker, which felt like the minimum bar.