Attachments and @mentions

Nomi works from what you show it: a photograph of the product, an ad whose look you want to borrow, a picture already in your Library. This page covers how to get one onto a message, which files are read, and what @ really offers.

Getting a picture onto a message

Press the paperclip beside the field. Select a reference opens a panel with a Library tab and a Saved tab, and Upload opens the file picker. Inside a session a third row holds the product, with its face and name, and opens the product list; on New task the product is picked on the Product pill beside the paperclip.

You can also drag a file onto the composer or paste one into the field. Everything attached shows as a thumbnail above the field, with an × to remove it.

What Nomi can actually read

The file picker accepts images, PDFs, Word documents, CSV, Excel and plain text. Only images are sent to Nomi. A document attaches and stays in your browser: its chip rides the message, and nothing is read out of it.

Of images, four formats travel: PNG, JPEG, GIF and WebP. Anything else attaches but does not go, and the message says how many did not come through.

How much fits

Four uploaded pictures per message, about 3.5MB together once Nomi has compressed them. Past either limit the extras drop and you are told, rather than the message failing.

Pictures named from the Library, from Saved ads or from a product have their own budget: four per message.

Marking up a picture before you send it

Click a picture you uploaded and it opens at full size. Under the close button is a pen: press it and a toolbar opens at the foot of the screen with two tools.

Comment is for saying something about one place. Click a spot or drag a box around it, then type what you want done there. Leaving the box empty removes it again, which is how a box you drew by mistake goes away.

Draw is a freehand line. Circle the thing you mean and say what you want in your message: the line points, the message asks. It carries no words of its own.

Undo takes back whichever of the two you added last. Press the pen again and the picture stops answering the pointer, with everything you drew still on it.

A picture with something on it wears a small pen badge on its thumbnail, with the count once there is more than one.

What Nomi sees

What you drew is painted onto a copy of the picture before the message is sent, because that is the only way an image model can be told where. Nomi is told the marks are annotations and never to draw them into the ad.

The instructions you wrote beside each marker travel with the painted picture. If the annotations cannot be applied, the original picture is sent and the message tells you that the annotations did not go through.

The copy is drawn in magenta whichever colour you picked. The colour on the swatch is for telling your own lines apart while you work.

Typing @

Every attached picture wears a small badge: @img1, @img2. Click it to rename the picture to a word you would actually use, like bottle. The number counts through the whole session, so a picture attached to your third message might be @img5, and the context shelf lists it under the same word.

Typing @ in the field offers those badges, then the pictures sent earlier in this session, newest first. It does not browse your Library and it does not offer products: it points at what this session already holds, so a name in your sentence always means a picture Nomi has. The one extra row is @nomi, which addresses the agent in a session the whole brand can read or one in a space with other people in it.

Typing /

Typing / opens Nomi's skills: /format, /copy, /research and the rest, plus any skill your brand wrote under Skills in the Agent's sidebar. Pick one and it sits in the sentence as a chip, and Nomi does that kind of work without deciding for itself.

/format takes a layout after the colon: pick it and the list moves on to the layouts on the shelf, and /format:text-on-skin tells Nomi to make the ad in the Text on Skin layout and no other. /save-format with an ad attached does the opposite: Nomi reads the layout off that ad and files it on the shelf, so later ads can be made in it.

/audit with a page's address audits that landing page as a phone and a desktop visitor. Without an address Nomi audits the landing of the product the session is about, and asks when there is none.

What a picture counts as

A picture you upload on its own is read as the thing you are selling, so its shape, packaging, colours and label text stay as they are. If the message also carries a product or a named reference, that upload is read as a reference instead: something for the ad to look like, never the thing it sells. Anything from the Library or Saved tab is always a reference.

The selected products provide context for the conversation. Their pictures are available for generation; they are not automatically attached to every message. Beside the paperclip a product's individual parts sit as a small stack, and pressing one writes that picture's name into your prompt.

The product selection is saved for the session, including when you remove every product. If saving fails, Nomi tells you. Your current selection still applies to the next message, but it may not survive a reload; select it again to retry saving.