Trust
Sources, confidence, asking first, and showing the work. 37 parts.
37 parts
Trust, 22 parts
- ClaimCheckA reply with the sentences no source backs up marked in place.
- SensitiveDataGuardWarns before a message with a card number, password or other private detail is sent.
- ConnectorPermissionOne connected app, and exactly what the AI can do with it.
- ContextPanelExperimentalEverything the AI can see right now, each item removable.
- AILabelA small lime tag that says content was made by AI and explains how. Once a person edits the content, it becomes an undo button.
- AIMarkA small four-point shape that tells people something was made by AI.
- AIBadgeA lime-tinted badge with a sparkle that marks content made by AI.
- AIDisclosureA small centered note reminding people to check the assistant's answers.
- ProvenancePopoverA label that says content was made with AI, and opens a short note on how.
- CitationMarkerA small numbered link that points to a source, placed after a claim in a reply.
- CitationHoverCardA small number in a sentence that shows where a fact came from.
- SourceCardA numbered link card for one source.
- SourceListThe numbered sources behind an answer, shown in two columns.
- ConfidenceIndicatorThree bars and a label that show how sure the assistant is about an answer.
- FeedbackBarThumbs up and down buttons, with a short list of reasons after a thumbs down.
- RefusalNoticeA calm notice for a request the assistant will not do, with an optional alternative.
- SummaryCardAn AI-written overview with its sources, a warning to double-check, and a way to give feedback.
- RevertToggleLets people flip between the AI's suggestion and their own edit, with a way back to the original.
- MemoryNoticeA short notice that says what was saved to memory, with Manage, Undo and close.
- MemoryListA list of what the assistant remembers. Each item can be edited or deleted. Two switches sit on top.
- PrivateChatBannerA notice that stays visible during a private chat, says how long a copy is kept, and has Exit.
- KeyTakeawayExperimentalA labeled box for the closing point of an answer.
Asking and showing work, 15 parts
- QuestionSetOne card that asks every missing fact at once, one question at a time, with a single final button.
- ActivityTraceThe agent's work shown step by step, each step tagged with who did it, then folded to one line.
- ReceiptCollapse a card the person already answered into one line. It opens again to show the full card.
- DraftPreviewA drafted message shown like an email, with headers and a body that opens up to the full text.
- CountdownA deadline as a large date, the days left in words, and an optional bar of time used.
- StatusLineOne quiet line of status for the top of a screen, such as Case, Flipkart, an amount and Waiting.
- MoneyAn amount that code worked out, in mono with tabular figures, in rupees by default.
- GlancePanelThe case at a glance: the deadline, the facts, the rule and the steps, kept in view beside a chat. GlanceSheet is the phone version.
- ScrollEdgeA scrolling area whose top and bottom edges blur gradually, so content softens near the edge instead of meeting a hard line.
- StatusTickerOne line that shows the real current step of an AI run, such as Reading 3 files, and keeps its width so nothing beside it moves.
- WorkingEdgeA calm moving ring in the AI colour around a surface, shown only while the AI is working on it.
- LiveLogA streaming, monospaced log of what an agent is doing, with pause and follow.
- JsonViewShows a tool call's input or output as a tree you can open level by level, with a copy button.
- InlineConfirmA button that asks "Delete this?" in its own place, with the safe choice focused, instead of opening a dialog.
- PromptIndexA short list of the person's own messages in a long thread, so they can jump back to one.