Aditya Umale's cleanup pass strips a named list of AI tells

Aditya Umale posted that 5 of the last 9 drafts his writing bot produced went out without him changing a word. The reason he gives is not a better prompt or a better model. It is a cleanup pass that runs on every draft and deletes a fixed list of patterns.

Aditya Umale

@AdityaUmale17

X

5 of the last 9 drafts my writing bot gave me went out without me changing a word

The list he published is specific, which is the useful part. It cuts the "it's not x, it's y" construction, one-line closers like "let that sink in", em dashes, forced lists of three, the phrases "here's the thing", "delve", "unlock" and "game-changer", a last line of "thoughts?" or "agree?", and emoji bullets with bold labels. He links the full list for anyone who wants to copy it.

What the number actually says

Five out of nine is one person counting his own drafts, on his own writing, with himself as the judge of what counts as shippable. There is no eval here and no before and after comparison against the same bot without the cleanup pass, so the honest read is that Umale finds the pass good enough that he stops editing most of the time. The only other datapoint in the thread is Milind S replying that it is really working, which is a reaction and not a measurement.

For people building on it

This is pattern matching on output, not better generation. A deny list of phrases removes the surface signals that a reader uses to spot a machine draft, and it does nothing to the structure, the argument or the thinness underneath them. If the draft had nothing to say, it will still have nothing to say, just without the em dashes.

The second thing worth noting is that a list like this ages. The tells on it are the tells of the models Umale is writing with now. Anyone copying it is inheriting a snapshot, and the honest version of this workflow is reading your own rejected drafts and adding whatever keeps showing up.