Matt Pocock’s /grill-me is built to interrogate a coding plan before you write any code — one contestable question at a time, until the design is sharp. I pointed it at a finished 99rabbits essay instead. It surprisingly worked.
The essay was about two patterns of how AI agents join human conversations — interceptors and demons. One question got the genre wrong, in a way that’s worth writing down.
The first question was: which single sentence would a reader quote afterwards and say this is what the post argued? It listed five candidates from the draft, ranked them, recommended the strongest one, and asked me to either endorse it or name a sixth.
My answer was that none of them was the thesis, because the piece doesn’t have one. It’s observational. It’s not arguing for a prediction or prescribing a behavior. It’s saying: here are two structurally different patterns, here’s how to tell them apart, here are the consequences each one carries. The contribution is the distinction itself, not a claim about which side wins.
The skill accepted that immediately and re-grounded:
Observational essays are a legitimate form: they don’t argue, they distinguish. The contribution is the distinction itself. Think Christopher Alexander on patterns, or any good taxonomy essay — before this, you couldn’t cleanly separate X from Y; after, you can.
Then it asked a different question, scoped to the new genre: does the opening deliver the distinction sharply enough? Which is the right question for an observational essay and the wrong one for an argumentative one.
Once the skill knew it was grilling an observational essay, the rest landed cleanly.
Q2. Second paragraph said “two structurally different patterns are in play, and I suspect that difference is worth thinking about.” The skill called out — correctly — that “worth thinking about” is doing none of the work. It offered three sharper openings, each with a different promise to the reader: clarity, recognition, design leverage. I picked.
Q3 was visual symmetry. The Pattern B section had a diagram and a screenshot. The Pattern A section had nothing. The skill named both the case for leaving it (form matches content — Pattern A is invisible) and the case for fixing it, and recommended fixing. I agreed. Inserted the diagram.
Q4 raised the strongest counter-objection a sharp reader would make; I judged the existing essay already defused it and overruled. Fine — the skill surfaces contestable choices, it doesn’t make them.
Three questions produced direct edits, one I overruled. The first cost a cycle to genre-correct, then re-routed. For finished prose, treat the first question or two as orientation — the grilling tightens once the skill knows what it’s grilling.
If you write nonfiction prose, try /grill-me at your next finished draft. Tell it the genre first; the rest ports cleanly.