LLM cliché highlighter

Simon Willison got tired of reading the same tics over and over in AI-generated writing, so he did something about it. He had an AI coding assistant called Fable 5 vibe-code a browser tool that hunts down ten of the most common phrases that give away machine-written prose. Paste in a block of text, and the analyzer lights up patterns like "no X, no Y" constructions, the weirdly therapeutic "sit with that," and the smug "you already know" — the verbal tics that have become so common in LLM output that they're starting to read as tells rather than style.

The irony is not lost on anyone here: an AI tool built to expose the fingerprints of other AI tools, assembled by yet another AI tool. But that's sort of the point. Willison's frustration came from reading an article stuffed with phrases like "no fluff, no filler, no jargon," the kind of hollow rhythmic padding that language models seem to reach for by default when asked to sound punchy or authoritative.

The tool itself is refreshingly unflashy. It runs entirely client-side, no server round-trip needed, and uses localStorage so your settings stick around between sessions. You can toggle individual patterns on or off, get counts of how many times each cliché shows up, and jump straight to the offending sentences via a match list. It's built for skimming a wall of text and immediately seeing where the AI-ness is concentrated.

What makes this interesting isn't the coding complexity — it's a fairly small, single-purpose web app — but what it represents. As more writing on the internet gets touched by language models, in newsletters, marketing copy, LinkedIn posts, even journalism, tools like this become a kind of literacy aid. They help readers, and writers, recognize when prose has slipped into autopilot mode, echoing the same structural crutches that models default to when nobody's paying close attention to the actual sentences.