Deslop is a Claude skill designed to revise drafts that sound machine-generated while preserving their facts. Its rules draw on corpus-linguistics and psycholinguistics research comparing human and model writing, checked against three public datasets where people and models answered the same prompts. The project argues that effective editing should target measurable patterns in context, not simply swap out a list of supposedly “AI” words. Users choose a genre and reader, then the skill scans a draft for stylistic flags, model-specific habits, and length relative to the prompt.
The workflow revises from larger concerns to smaller ones: content, structure, sentence grammar and length, word choice, and punctuation. It then compares the revision with the original, flagging newly introduced stylistic tells and checking that facts such as numbers, code, links, negations, and qualifications were not lost. A sample client email is cut from 228 words to 132 while retaining its launch dates and explanation. The project reports that a second pass focused only on stylistic tells made responses worse in a blind test, losing all five comparisons, so it advises stopping once the text falls within a human range. Basic scanning and fact checks require only Python 3; optional tools add linguistic, semantic, and model-based analyses.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.