yomiyasu skill fixes AI-slop Japanese at the syntax level

The skill grew out of a specific frustration with existing anti-AI-slop tools. Their instructions often used the very metaphorical verbs they told writers to avoid, and LLMs imitate the style of the prompt itself. They also told models to shorten the most important sentence to evade a sentence-length detector, which produced abrupt punchlines like 'Order is everything.' In response, the author built a corpus of 160 documents across eight practical genres and wrote a linter with lookaround regular expressions that ignore legitimate phrases such as 'grab the initiative' or 'melt ice'.

The design rests on three choices. First, instead of telling a model to replace words, the skill walks it through identifying the actor, reducing the event to physical facts, restoring subject-object-verb order, and flattening decorative formatting. Second, it forbids inserting short sentences just to vary rhythm and turns noun-heavy phrases back into verbs. Third, the skill's own instruction file was checked by the linter until it scored zero deductions, so the prompt itself serves as a model of the target style. A bundled Python script scores documents out of 100, deducting points for metaphorical verbs, decorative symbols, excessive bold and noun pile-ups.

Whether this approach travels beyond the author's own corpus is the open question. The boundary and human-written tests showed no false positives after the lookaround rewrite, but those 64 documents were assembled by the author. The test for teams is whether the rules survive contact with their own idioms, and the MIT license makes that easy to check. If it holds, the payoff is smaller: readers spend less time reconstructing who did what from context.