A classifier with no text generation pushes the runtime check out of the prompt and into code

On its way to a label, the old prompt was settling four separate questions: whether there is a feature-length film in the listing, whether there is more than one, whether this is television, and whether a film is actually screened or only discussed ^[10]. Those four are now asked directly, and they land exactly where Jev's documentation says it is weakest, on counting and runtime arithmetic, where asking each one narrowly is the recommended handling ^[12].

The reason the category descriptions had to be rewritten is in the docs. Jev "answers the question you wrote, not the one you meant", with scoping words and negations read at face value ^[13]. Several of the ten categories are defined mostly by what they exclude: a concert film belongs in `movie` and not `music`, and a film with a Q&A afterwards is still a `movie` ^[14]. A language model usually infers that from a loosely worded description, and a buried "not a comedy film screening" clause is less reliable here than the same exclusion given a field of its own ^[15]. Each option now carries an explicit `excludes` list, and the documented advice is to reach for it only once two options are similar enough that the model keeps confusing them ^[16].

Numbers go out of the request entirely. Jev is documented as not being a calculator, and it does better on semantic language than on numeric formats ^[17]. So instead of handing over `Duration: 73 minutes` and hoping, the maintainer does the comparison in his own code and passes the result as a fact, worded deliberately without naming a category so that it states the relationship ^[18].

Some of the daily volume is just title drift. The post's example response classifies "Star Wars: Episode VI Return of the Jedi (1983)" as a `movie` with a confidence of 9 ^[6]. That listing only needed categorising because The Movie DB calls the film "Return of the Jedi" ^[7].

The write-up documents the redesign and does not include an accuracy or cost comparison against the prompt it would replace ^[24]. That matters for anyone weighing the same swap, because the incumbent has a year of hardening behind it: at 450 to 600 listings a day, roughly 164,000 to 219,000 decisions ^[4][23]. Against that, the experiment cost an evening ^[9].

Some of the residual error will stay put whichever engine wins. The listing that opens the post is the quarterly meeting of the UK Buster Keaton Society at the Cinema Museum in Kennington, running 4pm to 7pm, with members requesting beloved gems to watch together ^[20]. Reviewing that batch by hand, the maintainer marked one listing as `multiple-movies` and gave up on the next, recording it as having no single right answer ^[21]. "Even for a human this is a hard problem to solve," he wrote ^[22].