Post

FYI: Your cliché highlighter got mentioned here ->

I'd pay $3/mo for this to be a browser extension.

If you're implying that somehow LLM systems will be tweaked to not use the AI clichés, that's more or less impossible due to how LLMs work.

To be Honest Neil, its a bit a futile excersize ;) . You can so easily add an editorial step in yr LLM request, certainly if you use an API direct of the LLM, it doesnt even use extra tokens... Ergo: ask the AI-LLM to check if it wrote as an LLM and correct it, or do a 2 way if you want. 🤷‍♂️🤷‍♂️

Not sure if you want pull requests, but I sent you one that adds another rule from the wikipedia article. github.com/simonw/tools...

Pattern lists age in one direction: they only ever catch what already got named. Worth logging which of the 38 never fires — those are the ones the models moved past.

Hi Simon. English prof here. I like your tool, but in a test it twice identified a "colon into a triple", though both had 4 items, misidentified a list of survey questions as rhetorical questions, plus other issues. I have thoughts about improvements, but as a stranger on the internet... ¯\_(ツ)_/¯

Argh

I don't know if this is specific for text generated in Italian, but I've been noticing that things are not only "real" but also "true", "actual", "concrete" (not the building material)

A cliche highlighter is useful because it makes style review concrete. Once the mush has names, teams can stop arguing vibes and fix the sentence.

22 patterns of Claude categorizes by Ivo open.substack.com/pub/linkandt...

This is great. How easy is obfuscation? Has anyone tried feeding these 38 patterns into a system prompt with a "never do these things"?

*eye twitching while reading that passage

Interesting, I've been using some (not all) of these patterns in my own writing for a long time. A mix of persuasive writing and marketing speak. I wonder if they've always been annoying or if it's just particularly bad when you use one (or multiple!) in every sentence 😅

The tells and the checklist are the same document. That list is also what you'd hand a rewriter told to strip them, so anything that has been through a de-slopping pass comes out clean too. Which leaves clean meaning either a human or a careful edit. Still catches raw output.

Gemini: "You hit the nail on the head!" is one of my favorites. I had a few hundred tells I collected in a few hours using searches on Google. Think there was a reddit that was about nothing else

I am gonna put my human writing through this occasionally to make sure I don’t slip into cliché-speak.

Adjacently, If you haven't tried pangram.com, you might find it interesting/useful.

I can't believe we still need to tell the LLMs to use sentence case

This is silly. LLMs are homogenizing our languages and flattening the diversity of expression through "correctness" of grammar. Don't let the AI edit your personality and style out of your writing.

The ones that survive a human edit pass are structural, not vocabulary. Nobody defends 'delve'. But 'To be clear' / 'I'm not saying' / 'Don't get me wrong', defending against an objection nobody raised, reads as ordinary hedging, so it stays in. Cheap to match on, too.

Reading this as rap lyrics ...