Hey again, it's Hasen.
Renaming files with Cowork is simple on the surface. Point it at a folder, write a prompt, and it runs.
What I hadn't thought through was what Cowork can actually see when it opens a file, and whether I need to set naming rules.
The Convention I Skipped
I started renaming invoices and receipts with no naming convention in place. Cowork still produced usable names, but a different format each time. Sometimes its date first, other times its doc type first. This depended on how I wrote the prompt that day.
The problem showed up later, when I needed to find one specific invoice and had no idea what pattern to search for.
Two File Types, Two Different Problems
The convention I eventually built covers invoices and receipts. Both categories need to be findable fast, at tax time, when tracking expenses, or when matching a receipt back to an order.
Product images and video haven't gotten there yet, and this is for a different reason.
We shoot everything for a listing in one session, so images and video land in the same folder. Early on, Cowork couldn't open an image and see what was actually in it, so naming relied on filename or metadata. However, a file called IMG_4821.jpg said nothing about what the underlying image is, not to Cowork nor to me.
Video compounded the problem, since Cowork couldn't view video files at all. That meant every mixed folder from a shoot had to be split into images and clips manually before any renaming prompt could start. That splitting is still the reality today, and it's part of why a convention for images hasn't been built out yet.
Why Some of This Fixed Itself and Some Didn't
The naming inconsistency was not a Cowork problem. Cowork does what the prompt tells it to do. No fixed rule in the prompt meant no fixed pattern in the output. That part was on me from the start.
The image problem was different. It was a real limitation, not a missing rule. Cowork used to guess a file's contents from its name or metadata, because it had no way to open an image and look at it. That's changed. It can now open an image file directly and see what's actually in it, the same way a person would.
That vision support isn't universal across every image format. JPEG, PNG, and GIF get read directly. Formats outside that set, TIFF, BMP, or HEIC straight off an iPhone, fall back to the old behavior: filename and metadata only, no real look at the image itself.
Video sits outside all of this. Cowork can’t view video files. There's no evidence of it opening a clip and watching it frame by frame the way it now does with a static image. The only workaround is pulling still frames out of the video and treating those as images.
That's the distinction worth holding onto. A missing convention is something you control. A missing capability isn't, and no amount of prompt tweaking gets around it.
The Convention, Ready to Steal
Here's what I built for invoices, receipts, and bills. Same rules, same reasoning. You can save it as a Cowork project or a skill.
The pattern:
YYYYMMDD_DocType_Party_RefNumber.ext
DocType is a fixed set of codes: INV for invoices, RCPT for receipts, BILL for utility or subscription bills, QUOTE, STMT. Five categories because forcing everything into three buckets loses information I actually need later.
Party is the supplier, shortened to a consistent slug. Red Shed becomes redshed, NZ Post becomes nzpost, always the same form for the same supplier. Consistency here matters more than cleverness, since the whole point is being able to search for one supplier and get every file back.
RefNumber comes from the document itself, the invoice or order number, padded to four digits if it's short.
Two examples from actual files:
20260610_INV_redshed_ORD88210.pdf
20260415_BILL_spark_INV0042.pdf
What happens when a field can't be extracted? A missing date becomes the literal placeholder YYYYMMDD, a missing supplier becomes UNKNOWN-SUPPLIER, and every placeholder gets flagged in a summary log for me to check manually. That log is the difference between trusting a batch rename and having to re-open every file to verify it.
This convention only covers invoices, receipts, and bills, not product images. Invoices had an obvious structure to design around, date, type, party, reference, all sitting right there on the document. Product images don't offer the same natural fields, so I haven't started building that system out.
The Takeaway
Cowork will rename files exactly as well as the rules you give it, and as well as its current ability to see the file. Fix the first with a written convention. Know the limits of the second.
Next time you hand Cowork a batch of files, write the naming pattern down first, even three lines in a project instruction. Don't let the prompt be the only place the rule exists.
Bonus Tip
For video, don’t try to make Cowork do a job it isn't built for. Claude doesn't have native video vision, and pulling still frames to fake it is slow and unreliable.
Gemini does this natively. Google's own documentation describes the model processing both audio and visual streams of a video directly, including describing and extracting information from clips up to 90 minutes. So, if you're generating overlay texts or captions for a reel, that's the tool for it.
Coming Up
There's no naming convention for product images yet. Next issue, I'll get into why images resist the same clean structure invoices have, and what I'm testing to build one.
Until next time,
Hasen
Infinite Output — Practical AI for solo founders.
