Approach One: Write Rules, Let the Tool Execute
Tools like Hazel represent this camp: you set conditions — filename contains a keyword, extension matches, source folder, how long since it was modified — and the tool moves, renames, or tags whatever matches. Once configured, execution is reliable and completely transparent; when something goes wrong, it's easy to trace back to which rule caused it.
The tradeoff is that rules can only catch patterns you thought of in advance. Invoice filenames vary wildly, and it's genuinely hard to write a rule that captures "this is an invoice" as a concept — at best you can do rough filtering by source folder or fixed keywords. Change how files get named (a new invoicing system, a new download source) and the rule quietly stops working, and you have to go back and patch it. The whole setup gets more complex the longer you use it, and maintenance cost keeps climbing.
Approach Two: Read the Content, Decide for Itself
Filewise takes the other path: instead of relying on filename or source, it actually opens the file and reads what's inside — PDF text directly, screenshots and scans through Apple Vision OCR — to work out whether it's an invoice or a contract, the amount, who issued it, the date, and generates a naming and filing proposal from that. However messy the original filename is, it doesn't matter, because the filename was never the input in the first place.
Take the same invoice: the best a rule engine can do is "move every PDF in Downloads to some folder" — it has no answer for what the PDF actually is or what it should be called. Filewise reads the amount, the issuer, and the date, and produces something concrete like "2026-06-06_IKEA_receipt_$180.pdf". That's the practical gap between the two approaches, on the exact same file.
Content Recognition Has Real Limits Too
Honestly, this approach has its own weak points. Poor scan quality, handwriting, and complex layouts all drag down OCR and model accuracy, and a purely local model's understanding genuinely lags the strongest cloud models. Filewise's answer isn't to push through and guess anyway — it sets a confidence threshold: strong signal gets auto-named, uncertain cases get queued for you to confirm, and anything it can't extract structured information from just keeps its original name rather than being guessed at. Recognition here isn't "fully automatic, no questions asked" — it's "automatic where it can be, asking you where it can't."
It's Not Actually Either/Or
If your organizing need is purely structural — "delete all .dmg files after 30 days", "move everything in this folder to the archive drive" — a rule engine is still the more direct, more predictable tool, and it doesn't need to understand content to get it right. But anything that hinges on "what is this file, actually, and what should it be called" needs understanding content, and that's where content-aware organizing saves far more effort than piling up rules — especially for a Downloads folder with messy sources and no naming pattern at all. See a concrete case in how to fix a messy Downloads folder automatically, and if you're more worried about a tool getting it wrong, read this.
