Some things need a screen

You have forty product photographs and space for six. An AI assistant can suggest a selection and explain its choices. To judge whether those six belong together, you need to see them together. The description leaves out their appearance.

Conversation gets you started without making you learn every control. Once the images are in front of you, the next step might be a tighter crop or a comparison between two options. You may want to hold five photographs in place and replace the sixth.

The interface has to support those decisions as well as the original request. What stays visible? How do you compare an earlier version? Can you change one image without asking the assistant to rebuild the whole selection?

Make the result visible

Ben Shneiderman described visible objects and small, immediately apparent changes in his 1983 paper on direct manipulation¹. His examples included text editors and spreadsheets. A crop tool offers the same short distance between acting and seeing the result: you can judge the composition while moving the boundary.

A text instruction can be the quickest way to crop a batch: “Make all of these square and keep the products centered.” After seeing the result, you might adjust one image by hand. It takes little effort to drag a boundary a few pixels; describing that correction could take longer than making it.

In Magic Ink², Bret Victor examines how the arrangement of information helps people answer questions. The essay is a design argument, not an experimental comparison of interfaces. It suggests a question worth asking of an AI product: what can someone learn from the result before typing another message?

A contact sheet might show that five photographs use a pale background and the sixth uses a dark one. Seeing them at the same size makes the difference apparent. You can decide whether it gives the collection variety or feels out of place. The assistant can explain its choice alongside the images.

Keep a stable working surface

I would keep the chosen six together, with alternatives nearby. You could replace one, compare it with the previous choice or ask for more options with similar lighting. The other five should stay where you left them. If the whole arrangement moves after every request, you have to find your place again before judging the change.

Eric Horvitz's 1999 work on mixed-initiative interfaces³ explored how automated assistance could coexist with direct manipulation. That question remains relevant when an assistant can find images, suggest edits and write explanations. Each contribution needs a place in the task that you can inspect and change.

The MCP Apps extension announced in January 2026⁴ lets tools present interactive interfaces inside a conversation. A tool can return a form or visualization that a person can operate, with interactions communicated back to the assistant.

Anthropic's June 2025 announcement of AI-powered artifacts⁵ described interactive applications built and used alongside a conversation. Both approaches give designers a way to put controls near the discussion. They still have to choose which controls the task needs and which results should remain on screen.

Share one version of the work

The assistant and the contact sheet need to use the same selection. Remove an image with a click, and the assistant should know it is gone. Ask for a replacement in the conversation, and the sheet should update. Otherwise, you can end up discussing a collection that no longer matches the screen.

That requires some care in the underlying data. Each image needs an identity that survives sorting. A crop must belong to a specific version. An edit needs a status so the interface can show whether it is running, finished or failed. A request such as “use the previous one” depends on a record of what changed, not just a conversation that mentions it.

The contact sheet also needs to work without a pointer or sight. People should be able to select images from the keyboard, hear which ones are selected through a screen reader and get useful image descriptions. Text or voice commands may be the preferred way to make changes. Those interactions need access to the same images, versions and selection state.

To test the design, ask someone to compare two possible selections. Watch what they keep beside each other and what they ask the assistant to explain. If they keep returning to an earlier image, it probably needs a place in the comparison. If they struggle to describe a crop, a direct control may help.

Choosing the six photographs is only part of this task. You also need to see why they work together, adjust the one that feels wrong and know which version will be used. The screen gives those decisions somewhere to happen.