Why the humanizer cannot be turned off
The last stage of every WriterDean pipeline is a lint pass that most tools would make optional. Here is why we made it mandatory.
Every content tool we looked at before building WriterDean had the same shape. You type a brief, a model writes a draft, and then it is your problem. Some added a tone slider. A few added a "make it sound human" button. None of them told you what was wrong with the draft in the first place.
The humanizer stage exists because machine writing has recognizable habits. Readers may not name them, but they feel them, and the feeling is "this was not written for me." The habits are consistent enough to check for. That is the whole idea.
What the lint actually checks
The humanizer skill is a list of patterns with examples of the pattern and examples of the fix. WriterDean turns that list into checks that run on every draft. A few of the ones that fire most often:
- Groups of three where the writer reached for a third item to make the rhythm land.
- Closing sentences that restate the paragraph in a tidier voice.
- Hedges stacked on hedges, so a sentence says almost nothing with confidence.
- A negation followed by a correction, used as a way to sound thoughtful.
- Setup phrases like "here is the thing" that delay the point.
Each check produces a finding with the sentence, the pattern, and a suggested rewrite. Some findings are scored. A few are hard fails.
Why hard fails
A score is easy to ignore. If the humanizer only produced a number, the number would drift down over time as people got busy, and nobody would notice until a reader did.
Hard fails stop a draft from moving to the approved state. You can still read it, edit it, and export it as a draft. You cannot publish it through WriterDean until the fail is cleared. The fix is usually a sentence or two, and the editor shows you exactly where.
Why you cannot switch it off
We considered a workspace setting. The argument for it is obvious: some teams have their own editors and want the raw draft.
The argument against is that the raw draft is not better. Every time we compared a draft before and after the humanizer pass with people who did not know which was which, they preferred the one that had been through the pass. The stage costs a few cents per article and about twenty seconds. There was no case where removing it made the output better, so we removed the option instead.
If you want to see what it changed, open the trace. The humanizer step lists every finding it acted on, and you can diff the input and output of that step alone.
What it does not do
It does not rewrite your voice. Brand context runs first and the humanizer respects it. If your voice rules say short sentences and dry humor, the humanizer will not add warmth.
It does not check facts. That is a separate gate, and it runs earlier.
It does not guarantee a detector score. We do not build for detectors. We build for the person reading the page, and the patterns we remove are the ones that make that person stop trusting the text.