The writing style of the guidelines I make the AI read had transferred into the text the AI writes

Summary

On August 12, 2026, I sent back all four articles that had come to the point of publication. I sent them back because they were hard to read. Before this send-back I had already pointed out the same kind of breakdown in style four times, and each time I added a rule or a machine check to the writing guidelines the AI reads. The breakdown still did not stop. This article is the record of how I stopped it. I gave up on adding rules, looked into the cause, and rewrote the body of the guidelines themselves.

The cause lay in the guidelines, and it fell into two layers of different kinds. The first layer was a contradiction between the contents of the rules: a rule that told the writer to replace standard kanji compounds with easier native Japanese words conflicted with another rule that forbids exactly that replacement. In the second layer, the body of the guidelines that states the rules was itself written in the broken style those guidelines are supposed to forbid. I fixed both layers together on the same day, so I have not been able to measure separately how much each fix contributed. Even so, I judge that as long as I have the AI read more than 160,000 characters of guidelines every time, that style transfers into the text the AI writes.

In the body I first describe how the breakdown in style went on even after four rounds of added rules, and set out in a table the two layers of defects I found in the guidelines. Next I describe how I rewrote the 16 core files without changing their meaning. I split the revision and the verification between separate agents. I also describe the 14 discrepancies the verifying agent detected. I judge the effect of the fix by two measures. One is how the findings on the four articles fell. The other is a reproduction test in which an AI in a new session with no context reads only the revised guidelines. Finally I set out the part that still cannot be called converged, and why verifying this kind of fix is difficult. In the closing section I introduce four outside papers on the effect input has on the style of output, with the bibliographic details I checked today. I mark which parts rest on objective grounds from outside research and which are measurements from my own operation.

What you can take away

For anyone who has an AI read guidelines and produce writing or other work from them, this article explains the following four methods in a form you can carry into your own operation.

• Apply to the body of the guidelines the AI reads the same checks you apply to its output. Count whether the body that states the rules still holds the breakdowns those guidelines forbid.
• Write into the guidelines themselves a rule that the style of the guidelines is not a model for the writer.
• In a rewrite that touches only style, split the agent that makes the revision and the agent that verifies that the meaning of the rules has not changed, and run them on different AI models. Whether the meaning got changed cannot be detected by the agent that made the revision.
• Measure whether the fix works with a reproduction test. Have an AI in a new session with no context read only the revised guidelines and write. Have a different AI judge whether the patterns of breakdown appear.

This article is a consideration based on the primary records of this site’s own operation.