Teaching Machines to Write Like People: A Corpus Study of Claudish and a Register Intervention

We report a corpus study of Claudish, the register that a large language model coding assistant produces by default, togetherwith a prompt based intervention intended to constrain it. Analysing 900 assistant messages, totalling 93,145 words of prose drawn from37 days of production coding sessions, we find that the register is not a fixed vocabulary but a generative system. Of 1,074 distincthyphenated compounds, 68 percent occur exactly once, produced by approximately fifteen recurring affix templates applied to whatevernoun is locally available. We identify seven classes of stylistic flourish and six recurrent sentence frames that together produce a sustainedeffect of adjudication. Because the register is generated rather than stored, prohibiting observed terms addresses only a minority of cases.We argue that the phenomenon is best understood as a failure of audience design, and we describe a register specification that reduces,without eliminating, the effect. We note that the model reproduced the register while documenting it, and that naming the pattern doesnot suppress it. The specification and the analysis procedure are provided for replication.