Tamirlan Askar · Product Design Lead
Taking Alice's first model picker from mid-project to launch
Summary
- Context
- Alice, Yandex's ChatGPT-class assistant. I design its core chat UX and took Expert mode over mid-project, through the September 2026 launch.
- Problem
- The chat was built for one fast model. Expert added a slower, smarter one, with no way to pick it, read its long reasoning or leave a long wait.
- What I did
- Two named options, Auto and Expert, instead of a "smartness" slider
- One answer with its reasoning folded away, instead of a stack of partial answers
- An "Answer faster" exit right under the reasoning, checked in a first-click test
- Result
- About 30% more people who open the picker now choose Expert (16% → 21%). My exit placement won the test, 94% vs 84%, and was adopted.
Situation
Expert mode gave Alice, Yandex's ChatGPT-class assistant, a slower, smarter model, and with it the chat's first model picker. I design Alice's core chat UX, took Expert over mid-project and owned it to the September 2026 launch: I decided how people choose the model and read its reasoning, and tested where the exit from a long wait belongs.
Task
Alice's chat was built for one fast model. Expert changed four things, each needing a design decision:
| Expert… | …so the chat needed |
|---|---|
| is a second model | a way to choose, the first in Alice |
| answers, then keeps thinking | one readable answer, not a stack of partial ones |
| thinks for longer | a way back to the fast model mid-wait |
| takes many files at once | a composer (the message box) and messages that hold up to 20 files |
Action
Two named options, not a smartness slider
Why. I rejected a "less smart ↔ smarter" slider: it casts the fast model as the dumb end of a scale, when its strength is speed. Two models need two named options.
Result. A composer button opens a sheet with Auto (Alice picks the model) and Expert. I renamed the options (the old "Fast" became Auto, captioned "Alice decides whether you need a fast or an expert answer"), made the button show the current mode, and added a shimmer that highlights it. After these changes, the share of people who opened the sheet and chose Expert rose from 16% to 21%, about +30%.
The naming matters because most people never open the sheet: about 95% of Expert traffic arrives through Auto routing. Auto has to read as a real choice, not a hidden fallback. The separate Fast option my argument points to isn't agreed yet.
One answer, the path to it folded
Why. Without a fold, an Expert answer stacked up: chip, reasoning, partial answer, reasoning, partial answer… and the answer repeated itself. Hiding the partial answers was only safe if the final one stood alone.
Result. Each new answer folds everything before it, reasoning and earlier answers alike, into one line. Only the latest answer stays in view.
Put the exit under the reasoning
"Answer faster" hands a slow Expert request to the fast model. The question was where it goes.
Why. My hypothesis: people find the exit faster right under the reasoning than above the composer. I checked it with a first-click test on a desktop mock, ~120 respondents per variant:
| Task (paraphrased): you don't want to wait, speed up the answer | First-click success |
|---|---|
| A · above the composer | 84% |
| B · right under the reasoning block (mine) | 94% |
Result. +10 percentage points, significant at p<0.05. My variant was adopted; it isn't live yet.
Up to 20 files in one message
Why. Files are where Expert pays off most. Every file other than an image goes to Expert, and on files it leaves about a third fewer requests unsolved than the older models. But the composer and message cards were never designed for many files, so the case where Expert is strongest was the one the chat handled worst.
Result. After auditing ChatGPT, Claude and Grok, I designed the states Alice lacked (one doc, two docs, photo + doc, mixed) and the overflow:
| Where | With up to 20 files |
|---|---|
| Composer | one row, horizontal scroll |
| Sent message | a grid capped at 6, plus Expand, so one message can't take over the screen |
Result
- Sep 2026 Expert live: Alice's first model picker, which I owned to launch
- 84% vs 94% first-click success for "Answer faster", A vs B (mine, adopted)
- 1 answer not a stack of repeating partial answers; the reasoning folds away
- 2 options Auto and Expert, named, instead of a smartness slider
What I'd do differently
- Keep pushing for a separate Fast option, which isn't agreed yet. Speed is a strength of its own, yet Fast can't be picked directly.
- Judge the next iteration by behavior: abandon rate while Expert thinks, and repeat Expert use. Tests checked the pieces; those numbers would check the whole.