{"slug": "i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai", "title": "I let AI write this post. Then I deleted everything that made it sound like AI.", "summary": "A developer created Alchemy, a ruleset that removes common AI writing patterns from generated text. The tool bans overused words like 'delve' and 'robust', limits phrases such as 'in the ever-evolving landscape', and flags constructions like the rule-of-three flourish. Alchemy is available as an MIT-licensed Markdown file on GitHub.", "body_md": "A language model wrote this post. I directed it, but I didn't rewrite a single sentence by hand. Then I ran the draft through a small ruleset called Alchemy. If the writing below doesn't trip your \"this is AI\" alarm, that's the entire point, and I'll show you the machinery.\n\nStart with the alarm itself. You've felt it. A README opens with \"In today's fast-paced digital landscape.\" A reply leans on the em-dash three times in two sentences. Something *delves* into something else. The words are grammatical and the facts might even be right, but a part of your brain has already filed it under *generated* and dialed down how much it trusts you.\n\nThat reaction is real and it's getting sharper. So the question for anyone shipping AI-written text is no longer \"is it correct,\" it's \"does it read like a person bothered to write it.\"\n\nHere's a paragraph describing this tool, the way a model writes it with no guardrails:\n\nIn today's fast-paced digital landscape, writing has never been more important. Alchemy is a robust, powerful, and seamless solution that empowers developers to elevate their content and unlock the full potential of AI. By leveraging cutting-edge techniques, it delves into the intricacies of language to deliver a truly transformative experience — a game-changer for anyone looking to take their writing to the next level.\n\nForty-eight words and it tells you nothing. Now the same point after Alchemy:\n\nAlchemy is a list of rules you hand your AI. It deletes the patterns that give generated text away: the em-dashes, the \"not just X, but Y,\" the words like\n\ndelveandrobustthat nobody says out loud. What's left reads like a person wrote it.\n\nThe second one isn't smarter. It just stops performing.\n\nThe easy version of this post would list words I find annoying. Instead the ruleset is built on what people have actually measured.\n\nWikipedia's editors maintain a page called \"Signs of AI writing,\" a long catalogue of tells they use to clean up machine-generated articles. A 2025 study by Kobak and colleagues went through millions of biomedical papers and found specific words whose frequency jumped after ChatGPT shipped: *delve* showed up at roughly 28 times its old rate, *underscore* and *showcase* close behind. Pangram Labs did the same for phrases and found \"in the ever-evolving landscape\" running about eleven thousand times more often than humans ever wrote it.\n\nSo Alchemy bans the heavy hitters outright, puts a second tier on rations, and flags the constructions: the rule-of-three flourish, the recap conclusion that repeats the intro, the helpful-assistant outro.\n\nOne rule matters more than the lists, though. No single word proves anything. Real writers use em-dashes, *delve* is a fine word, and a humanizer that strips every flagged term leaves prose worse than it found it. The tell is *density* — a dozen of these tics stacked together over flat, evenly-weighted text. The rules weigh co-occurrence over any one hit, which is also why this post can use an em-dash without flinching. This is the second one. It's fine.\n\nThe whole product is one Markdown file your agent reads before it writes. Drop it in:\n\n```\nnpx @fernforge/alchemy init\n```\n\nThat writes `ALCHEMY.md`\n\ninto your project and links it from your `CLAUDE.md`\n\n, `AGENTS.md`\n\n, or `.cursorrules`\n\n. From then on the rules ride along whenever the agent writes a doc, a commit message, or a reply. You can also just read the file: it's MIT-licensed and lives at [github.com/fernforge/alchemy](https://github.com/fernforge/alchemy).\n\nIt won't make a model say anything true or interesting. That part is still on you. What it does is stop the writing from announcing where it came from, so the ideas get read on their own terms.\n\nWhich, given how this post was made, is a claim you're in a good position to judge.", "url": "https://wpnews.pro/news/i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai", "canonical_source": "https://dev.to/fernforge/i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai-28aj", "published_at": "2026-06-27 13:38:02+00:00", "updated_at": "2026-06-27 14:03:58.977269+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "developer-tools"], "entities": ["Alchemy", "FernForge", "GitHub", "Wikipedia", "Kobak", "Pangram Labs"], "alternates": {"html": "https://wpnews.pro/news/i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai", "markdown": "https://wpnews.pro/news/i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai.md", "text": "https://wpnews.pro/news/i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai.txt", "jsonld": "https://wpnews.pro/news/i-let-ai-write-this-post-then-i-deleted-everything-that-made-it-sound-like-ai.jsonld"}}