The writing quality system passed a draft clean. Every automated check came back green. Then I read it.
There was a paragraph listing out GitHub repos — not because the reader needed them, but to establish that I have a lot of them. A few lines later: a soft pivot toward paid work, dressed up as the natural next step in the argument. Neither paragraph set off any alarms because neither sounded like AI. The craft was fine. The slop was absent. What those paragraphs were actually doing — establishing credentials, nudging toward a sale — went undetected.
This is the specific failure the system was built to catch. The irony writes itself.
The gap makes sense once you see it. Phrase-level checks are good at catching how something sounds. They’re blind to what the content is doing. A paragraph can be grammatically clean, free of hedging, free of filler — and still be serving the writer instead of the reader. Those are different problems, and the system only solved one of them.
The fix I landed on is a paragraph-level intent test I’m calling the Positioning Energy Check. One question per paragraph: who does this serve? If the honest answer is “it makes me look good” or “it moves the reader toward buying something,” that’s a flag. Not an automatic cut — a credentials paragraph can earn its place, a soft close can be legitimate — but it has to pass a conscious “why is this here?” test, not just a “does this sound human?” test.
What I’m noticing is that the harder problem isn’t quality — it’s intent. You can build a system to recognize AI-flavored prose. It’s much harder to build one that notices when a writer is quietly serving themselves.
Still not sure if this check should run automatically or stay in a manual review step. The automatic version risks flagging legitimate confidence as bragging. The manual version requires remembering to ask. Neither feels right yet.