dsh-humanizer

AI-generated text carries a recognizable machine accent: filler words, formulaic closers, parallel structures nobody asked for. Rewriting sentence by sentence is slow, and stripping the tone often distorts the original meaning. dsh-humanizer moves the intervention earlier, so the style is in place before a single word is generated.

It does two things. First, it injects a positive writing rule set into the system prompt before generation: be specific, hold a point of view, vary sentence length, drop filler, skip the grand conclusion, no emoji. Alongside that sit three non-negotiable fact rules: never fabricate, never alter facts, and when tone and fact conflict, the fact wins. Numbers, dates, names, and quotations are preserved; if stripping the accent would distort meaning, the accent stays. Second, it ships a detect_ai_flavor tool that scans text against 22 enumerated patterns (10 Chinese, 12 English) using pure code. It returns a 0-100 score, a human / mild / ai verdict, and the matched patterns with samples. Zero tokens consumed, fully reproducible.

Built for anyone who wants a human touch in the output: content editors, technical writers, developers producing copy at scale. You do not need to proofread line by line; set domain and voice in your profile and the style constraint rides along on every generation. To gauge how much a passage reads as synthetic, just ask the model to run the detector and you get a score and a pattern list. It is a ruler, not a judge: literal matching, density-weighted, good enough to calibrate your instincts without rewriting on your behalf.