Same assistant. Different posture.

01 / HOME

New + recent threads

A Messages-like list. Shortcuts stay compact. No composer until New.

02 / FOCUS

@ a person or chat

Inviting someone into the research room—avatars, chips, scoped sources.

03 / ANSWER

Cited Claire cards

Answers are objects: sources, follow-ups, never a sourceless paragraph.

04 / LOOPS

Open loops & tone

Structured cards you can track. Observation separated from inference.

05 / AMBIENT

Quiet in-chat Claire

One ribbon. Inline promises. No reply strip until you ask.

06 / SHEET

Same thread, pre-scoped

The in-chat sheet is Ask Claire with Maya already focused.

07 / CATCH UP

Needs you vs FYI

The stale-chat moment: who is waiting, then one suggested next line.

08 / SEND

Rewrite the draft

Shorter, warmer, clearer—on the text you already typed.

What earns its place first.

SHIP FIRST
  1. Ask Claire home + cited answers + @ Already built. The UI is the gap. Highest unique value.
  2. On-demand drafts only Keep suggested replies. Stop generating on every inbound.
  3. Catch me up Especially stale and group chats. Cacheable. In-chat first.
  4. Open loops as cards + Track Connects Ask Claire to Promises. Retention, not novelty.
SHIP SECOND
  1. Rewrite the draft you typed Shorter / warmer / clearer. Tiny context, send-time magic.
  2. Cross-chat pull from inside a chat @Noah while talking to Maya. The thing no clone can do.
  3. Tone check on your outgoing draft Regret prevention. Observation, not a personality essay.
FINISH, DON’T EXPAND
  1. One ambient ribbon Keep Quick Context. Do not also launch a smart-card tray.
  2. Relationship memory as fuel It should feed drafts and tone, not be a destination.
  3. Not core: auto-send, plugins, docs Ask After research + draft already feels magical.

The rules behind the screens.

Claire never sends from Ask Claire

Research and drafts stay reversible. Use reply drops text into the composer. The person still hits send. Auto-reply that sends is not a launch AI surface.

Citations are required

Every answer names the messages it used. Sources open the chat with a highlight. Foreign sources say they came from another conversation.

Generate on intent

User-tapped actions only. Suggested replies generate when the sheet opens or when someone asks for help—not on every inbound message.

@ is scope, not syntax

Focusing a person should feel like inviting them into the room: avatars, lime chips, and an answer header that says who Claire is looking at.

Observation versus inference

Tone checks separate what the messages show from what Claire is guessing. No personality diagnosis. A matching next line is the payoff.

One ambient surface

In chat, keep the Quick Context ribbon and inline promises. Do not also launch a smart-card tray. Presence is quiet until asked.