Claude UX Teardown: Settings Before the First Word

Claude's new-chat page is mostly empty. A heading, one text field, and under the field a row of small controls: Project, Output, a model name with an effort level beside it, and a word, "Manual".

The incognito page

The capture was taken in Claude's Incognito mode. A ghost icon and the word "Incognito" sit in a thin bar at the top, with a close button on the right. The heading reads "You're incognito", and the composer's border is dashed rather than solid. Below the controls, in grey: "Incognito sessions stay out of your history, memory, and search, and aren't used to train Claude", followed by a Learn more link.

So the mode is announced three times: the top bar, the heading, and the sentence, which is the only one of the three that says what the mode does. The dashed border looks like a fourth cue, though nothing on the page says so.

Seven items behind the plus

Add files or photos comes first and is the only row with a keyboard shortcut (⌘U). Then a divider and five rows with chevrons: Skills, Connectors, Design system, Plugins, Devices. Each chevron promises a submenu, and the capture opens none of them. A final row, Research, has no chevron and stands alone after a second divider.

The menu follows progressive disclosure (NN/g): the composer shows a single plus, and the breadth sits one click down. Five of the seven entries, though, lead to a further level. Someone who wants to know what "Devices" means here has to open the menu, then the submenu, to find out.

Four models and an effort row

Clicking the model name lists four models, each with a short line about the work it suits. Fable 5.1, "For your toughest challenges". Opus 5.5, checked, "For complex work and everyday tasks". Sonnet 5.5, "Most efficient for simpler tasks". Haiku 4.5, "Fastest for quick answers".

They don't divide the work cleanly. Opus, the default in this capture, claims both "complex work" and "everyday tasks", which overlaps Fable at one end and Sonnet and Haiku at the other.

Under the models, separated by a rule, is Effort, set to Medium, with its own chevron, and below that More models. Effort lives inside the model selector, and in the chip row it appears as grey text right after the model name: "Opus 5.5 Medium".

Output is chosen before the prompt

Document, Presentation and Design each carry a Beta tag. Google Docs, Google Sheets and Google Slides follow with their own icons. After a divider, Artifact, with the grey note "Let Claude pick the format". No row has a check mark, and the chip itself still just says "Output", so from this screen it isn't possible to tell what happens if you choose nothing. Artifact's note suggests it's the hand-it-to-Claude option; the capture doesn't confirm whether it's also the default.

Labels on the chip row

"Project" and "Output" name the decision; you have to open them to see the current choice. "Opus 5.5 Medium" names the current choice, two of them in fact, and you have to open it to learn that Medium is an effort level. "Manual" reads like a current value too, but the screens never show what question it answers. The capture doesn't open it, and nothing nearby explains it.

Visibility of system status (NN/g) asks that people can see the state they're in. Here state is shown for the model and not shown for project and output.

On the last screen the prompt is typed: "Help me compare two onboarding concepts for a mobile budgeting app. Give me a concise decision matrix." The microphone has been replaced by an orange send arrow. Everything on the chip row is exactly as it was on the empty page, still reading "Project", "Output", "Opus 5.5 Medium", "Manual". At the moment of sending, the row still doesn't say which format the reply will come back in.

Putting values on the chips

Show the value on every chip: "No project" and "Output: Auto" (or whatever the real default is) cost a few characters and save two clicks of checking. Put a word in front of the effort level, so the chip reads "Opus 5.5 · Effort: Medium", and give "Manual" the same treatment so it says what it is manual about. In the model selector, rewrite the Opus line so it stops overlapping its neighbours, or mark it "Recommended" if that's what it is.

Take it further

This is the "can I see what I've already chosen" check from our UX Clarity framework, which we run on whole products in a Full UX Audit.

These five screens come from claude.ai/new on 28 September 2026, signed in to a personal account with Incognito switched on and the account sidebar collapsed, at 1440×900 in English. The prompt was typed and never sent, so the reply, any artifact it would have produced, and the Effort, Manual, Project and plus-menu submenus are all unseen here.

Sources: NN/g — Progressive Disclosure · NN/g — Visibility of System Status.

Does your own interface show what's currently set before people hit Send? Apply for a Full UX Audit →